Improving E-Commerce Sales Using Machine Learning

by

Shaheen Thobani

B.E. Industrial Engineering

Nagpur University (1988)

Submitted to the System Design and Management Program

in Partial Fulfillment of the Requirements for the Degree of

Master of Science in Engineering and Management

at the

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

February 2018

C 2018 Shaheen Thobani. All rights reserved

The author hereby grants to MIT permission to reproduce and to distribute publicly paper and electronic copies of
this thesis document in whole or in part in any medium now known or hereafter created.

Signature redacted
Signature of Author

System Design and Management Program

January 19, 2018

Certified by Signature redacted-

(i Michael Cusumano
Thesis Supervisor
Sloan Management Review Distinguished Professor of Management

Signature redacted
Accepted by
MASSACHUSTS IN UTE Joan Rubin
OF TECHNOLOGY. Executive Director, System Design & Management Program

JUN 2 0 2018
LIBRARIES
ARCHIVES
 77 Massachusetts Avenue
Cambridge, MA 02139

MITLibraries http://Iibraries.mit.edu/ask

DISCLAIMER NOTICE
Due to the condition of the original material, there are unavoidable
flaws in this reproduction. We have made every effort possible to
provide you with the best copy available.

Thank you.

The images contained in this document are of the

best quality available.
 Improving E-Commerce sales using Machine
Learning
by

Shaheen Thobani
Submitted to the System Design and Management Program
on January 15, 2018, in Partial Fulfillment of the Requirements for the Degree of

Master of Science in Engineering and Management

ABSTRACT

Trends show promising growth of the online e-Commerce industry. While the e-Commerce companies are
aggressively moving towards digital sales and marketing, the customers are being bombarded with frequent
and often irrelevant marketing communication from myriad sources. The thesis proposes understanding the
digital purchase journeys of the customers from the lenses of both sellers and customers to make online
sales and marketing efforts relevant and intelligent. The thesis applies the improved customer journey
framework to identify the needs of the customers and goals of the seller at various stages of customer
purchase journey. It discusses the need to take an integrated view of the purchase journey to improve the
customer experience at the journey level. It illustrates with an example how to design end-to-end journeys
- a starting point for consciously shaping the purchase journeys.
Larger companies are using Machine Learning to improve marketing technologies and processes to create
a competitive advantage and capture market share through digital presence. The thesis aims to understand
and illustrate the applications of Machine Learning to digital sales and marketing ecosystem for the e-
Commerce industry. It first understands the e-Commerce touchpoints using which customers interact with
the brands and delves deeper into the underlying technologies powering these touchpoints. Then it
illustrates and analyzes the application of Machine Learning to the e-Commerce website which includes
search, recommendation system, and Product Detail Page with an aim to improve conversion, and to the
advertising ecosystem which includes Data Management Platform and Demand Side Platform in order to
enable prospecting and customer targeting. The thesis also illustrates and proposes the use of a framework
called 'Machine Learning Canvas' to systematically apply Machine Learning to any system while keeping
value proposition for the business in the center.

Thesis Supervisor: Michael A. Cusumano

Title: Sloan Management Review Distinguished Professor of Management
MIT Sloan School of Management

2
 ACKNOWLEDGEMENTS

I am grateful to all the fantastic teachers I have met during my education at MIT. They have contributed
tremendously to my growth and learning. I am also grateful to all my schoolmates and friends for the time
I have spent with them as it helped me evolve intellectually and culturally.

First, I would like to thank my thesis advisor Professor Michael Cusumano for his advice, dependable
support, and patience during my thesis journey. In his classes and presence, I have learned more than
what can be taught.

I am grateful to Joan Rubin and William Foley for their constant support and guidance throughout the
program, and making the journey at MIT pleasant.

Finally, I am thankful to my family, my pillar of strength and support, for always being with me in all my
decisions.

3
 TABLE OF CONTENTS

ABSTRACT

ACKNOWLEDGEMENTS

TABLE OF CONTENTS

LIST OF FIGURES

1. Introduction
1.1 Motivation
1.2 Approach and Objective
1.3 Scope
1.4 Structure
2. Customer Purchase Decision Journey
2.1 Introduction
2.2 Evolution of the Customer Purchase Journey
2.3 Understanding the customers' purchase decision journey (CDJ) through the lenses of
sellers and customers
3. Customer Touchpoints
3.1 What are touchpoints?
3.2 Mapping touchpoints
3.3 Designing the journey
3.4 Considerations for designing the customer journey
3.5 Touchpoints - an integrated view
3.6 Important touchpoints for Ecommerce
3.6.1 Online Ads
3.6.2 E-commerce website
4. Technologies Underlying Touchpoints
4.1 Recommendation Systems
4.2 Applications of recommendation systems in e-Commerce website
4.3 Approaches to recommendation systems

4
 4.3.1 Collaborative filtering
4.3.2 Content-based filtering
4.3.3 Hybrid recommender system
4.4 E-Commerce Product Search
4.5 Advertising
4.5.1 Programmatic Advertising
4.5.2 Key players and components of the advertising ecosystem
4.5.3 Technologies enabling advertising
4.5.4 Programmatic advertising - a closer look
5. Applying Machine Learning
5.1 Introduction to Machine Learning
5.1.1 Description
5.1.2 Classification of ML
5.1.3 Choosing a model
5.2 Machine Learning Canvas
5.3 Application of ML to Recommendation Systems
5.4 Application of ML to e-Commerce Search
5.5 Application of ML to advertising ecosystem (DSPs) for prospecting
5.6 Application of ML to Propensity Modelling for customer targeting
5.7 Application of ML to Product Detail Page Optimization
6. Conclusion
6.1 Thesis summary
6.2 Challenges and opportunities for future work

5
 LIST OF FIGURES
Figure 1 Correlation of Retail e-commerce with smartphone penetration
Figure 2 Classic funnel metaphor of a linear customer journey Vs evolved non-linear journey
Figure 3 Evolved loop metaphor customer purchase journey
Figure 4 Classic Journey Vs Accelerated Loyalty Journey
Figure 5 Customer Journey Framework
Figure 6 Detailed TV purchase decision journey analysis from customer lens
Figure 7 Brand Touchpoint Wheel for all company touchpoints
Figure 8 Brand Touchpoint Wheel for customer touchpoints
Figure 9 Customer journey maps through touchpoints
Figure 10 Purchase Decision Journey of Persona 1
Figure 11 Purchase Decision Journey of Persona 2
Figure 12 Correlation of customer satisfaction and Willingness to recommend with touchpoints
Vs Journeys
Figure 13 Method for inbound marketing
Figure 14 Recommendations on the landing page of the e-Commerce website
Figure 15 Recommendations on the Product Detail Page
Figure 16 Feedback profile, a form of recommendation system on eBay
Figure 17 Steps for memory-based approach
Figure 18 Architecture of content-based system
Figure 19 Representation of Programmatic Advertising Ecosystem
Figure 20 Representation of Adobe's Data Management Platform
Figure 21 Programmatic transaction types
Figure 22 Representation of advertisement inventory transaction process
Figure 23 Real-Time Bidding process
Figure 24 ML Algorithm Cheat Sheet
Figure 25 Machine Learning Canvas
Figure 26 ML canvas applied to ML for recommendation system
Figure 27 User Probability Matrix
Figure 28 ML algorithms for recommendation systems
Figure 29 ML algorithms for e-Commerce search
Figure 30 ML algorithms for customer prospecting
Figure 31 Representation of Propensity Model
Figure 32 Product Detail Page

6
 1. Introduction

1.1 Motivation

Research forecasts that e-commerce will grow from approximately $2.8 trillion market - currently, 10
percent of the pie that represents $28 trillion retail markets worldwide to a $4 trillion market globally by
the year 2020. A study by Nielson shows that the retail e-commerce growth is strongly correlated with
smartphone penetration in most of the markets. Though the smartphone penetration does not completely
explain the growth of retail e-commerce, the trend does imply growth in online retail shopping [1]

a','a
34% UMJ!E6MN
16% 0M

.......W.... .EA....W0 .... 2%0 1

DA BVA 1 010 ODA 0%

KQ% 0% 02% M %%

u fIMmEiMiCA &OIMIm I4NEA a AS = nBWE

Figure 1, Correlation of Retail e-commerce with smartphone penetration [1]

A report published by Forrester observed and predicted that though the e-commerce market shows
promising future, the gap in the market is widening. [2] Some of the top players such as Amazon will
continue to aggressively gain market share and out-do many small e-commerce retailers. Among the losers
are mostly the smaller e-commerce retailers partly because they do not have more sophisticated ways of
reaching out to their potential customers. Larger firms have recently started adopting the advanced data
analytics, Machine Learning, and Artificial Intelligence in order to make their customer reach accurate.

On one hand, the e-commerce retailers are increasing the possibilities of reaching out to potential customers,
on the other end customers are getting bombarded with advertisements and recommendations through their
smart devices. Though the sellers now have so much more data about the customers, not knowing how to
use that data is resulting in innumerable and unsophisticated sales and marketing efforts. This is making
consumers insensitive and ignorant due to their irrelevance of content, time and place, to the extent of
having them look out for tools such as Ad-blockers. Using the predictive algorithms of machine learning
can let companies identify their potential customers and target them, improving their return on investment
on their marketing and sales efforts.

7
In order to improve their sales, online retailers do not only need to target right customer, they also need to
target them at the right time and during the right stage in their purchase decision journey and on the right
place (device or location) where the probability of catching their eyeballs is highest. In addition to
improving the customer reach, the companies also need to leverage analytics methods such as Machine
Learning to make their processes efficient.

1.2 Approach and Objective

The objective of this thesis is to systematically study the application of machine learning in the online sales
process with a goal to improve it. This thesis also attempts to understand the progress that has been made
in the application of Machine Learning in the space of retail e-commerce. The thesis approaches the e-
commerce sales using the framework of Customer Purchase Journey and then further diving deeper by
looking at the level of touchpoints used by customers to navigate through this journey. Then it attempts to
understand the technologies used to run these touchpoints and applications of Machine Learning to these
technologies.

1.3 Scope

The thesis limits the scope of its application and relevance to commodity products in retail e-commerce
including apparels, FMCG, and electronic products. It is important to note that the customer purchase
journeys for products such as automobiles, real estate properties, music subscriptions and all services is
different than for the commodities. Further, the touchpoints considered in this thesis are online touchpoints,
whereas the touchpoints used during the purchase journeys of above-mentioned products and most services
are mix of online touchpoints (used for research and awareness) and offline touchpoints such as customer
service phone calls and retail stores (for other activities of purchase decision journey including purchase).

1.4 Structure

This thesis has four chapters. The first chapter starts by describing the evolution of customer's purchase
decision journey from the metaphor of 'funnel' which implies more of a linear approach to understanding
customers' actions and steps from the motivation of purchase to the moment of purchase and later to
metaphor of 'loops' that takes into consideration the non-linear nature of the journeys today as a result of
connection between various stages of the journey and the stage of purchase. Further, the chapter presents a
framework of purchase journey from the seller and customer point of view, and delves deeper into their
objectives at each stage, and ends with an example of the application of the similar framework to explore
the journey of a customer who wants to purchase a TV.

8
The second chapter defines 'touchpoint' - the front end that the customers interact with and briefly
describes touchpoint mapping. Next, it recommends the process and considerations for designing customers
journeys with an understanding of touchpoints. The last section of the chapter describes most major online
touchpoints for e-Commerce customers which are broadly various types of online advertisements and the
e-Commerce website.

The third chapter delves into the backend technology of the touchpoints covered in the second chapter.
Before applying Machine Learning, it is important to understand the process and system that manages the
data, so the chapter starts with understanding the complex data management platform end-to-end starting
from the sources and ways of data collection to -the data delivery to the publishers of the online
advertisements. Further, it presents the study of advertisement technology and finally ends with the brief
study of the technology behind systems of major touchpoints on the e-Commerce website.

The fourth and final chapter presents and analyzes the current Machine Learning application in e-
Commerce website Search and Product Description Page, the Recommender System and the Ad targeting.

9
 2. Customer Purchase Decision Journey

2.1 Introduction

This chapter discussed the evolution of the Customer Purchase Decision Journey (CDJ) and decribes the
framework that can be used to understand the journey from the lenses of the customers as well as the sellers.

CDJ is the model that describes how customers make purchase decisions that start with the day to day
influences of brands and go up to the point when customers purchase and reconsider or recommend the
product they bought. [3] The survey by Salesforce shows that 86% of senior level marketers believe that it
is very important to have a cohesive customer journey. This journey needs to be consistent and make the
process as painless as possible for the customer. [4] Google provides their analytics solutions at not only
touchpoint levels but journey level. They assert 'It's important to understandthe entire customer journey
so you can measure all of the elements that contribute to your campaigns, attribute the right value to them,
and adjust your marketing budgets where appropriate.'[5]

2.2 Evolution of the Customer Purchase Journey

Traditionally, the consumer decision journey has been understood through the metaphor of a "funnel"
(Figure 2 Left) - the consumers start with a large number of options at the top of the funnel for the products
they want to purchase and systematically narrow down the choices to purchase one. This notion was true
until nearly a decade back when consumers were introduced to options through very few channels or
through their visit to the brick and mortar stores after which they evaluated and narrowed the number of
options. Today, the consumers are continually exposed to various products through multiple channels that
leave the impressions of products and brands in their minds. The multiplication of products, fragmentation
of the marketing channels and evolution of media have led to the reduction of the number of brands
customers remember and start their journey with. [6] When the consumers begin to consider buying an
item, these accumulated impressions inform the potential options they may start with. The number of
options might increase rather than decrease in the next step of the purchase journey when consumers reach
out to digital channels to evaluate their decision, or when consumers go to on the product page of an e-
Commerce website and see recommendations and options for a product they set out to purchase.

The analogy of funnel was appropriate until a decade back because then the purchase journeys were
relatively linear. However today, the increase in the types of digital channels that potential consumers use
and are exposed to has resulted in non-linear purchase decision journeys. The customers might interact with
new brands at any stage of their purchase journeys. In some cases, brands that were initially not being
considered by the consumer may enter into customer's decision journey in the later phases, of the purchase

10
 journey, sometimes resulting in customers dropping the brand they were most considering earlier. This non-
linear nature of the journey provides opportunities for brands to enter into the customer's decision
consideration when it is most efficient and makes the most impact on the customer [6] The customers'
decision journey evolved and became more complex than a linear funnel as sellers engage with their multi-
touchpoint customers with a multichannel approach [7] (Figure 2 Right)

Awaenms oamto Purchs*

3 DS

moum

U Web EConmmrebsa, U Eta Use N rs u sceI 3 Emad U Dwiz

Figure 2, Classic funnel metaphor of a linear customer journey Vs evolved non-linear journey [7]
As the conversation between sellers and the consumers started becoming two-way from one-way, the
metaphor of the consumers' purchase journey evolved from a funnel to a circular loop (Figure 3) [6]. The
loop implies that many sellers can successfully convert buyers to become long-term loyal customers giving
rise to a loyalty cycle between the 'Trigger' stage, a time when consumers discover their intent to purchase
an item and 'Purchase' stage, time when they purchase the same brand. For example, Sephora, a French
brand of high-end cosmetic products has a very strong loyalty program and it creatively stays in touch with
its customers post their purchases through Sephora mobile app by sending notifications and reminders to
reorder to a particular Sephora product their customers precisely when they are about to run out of it. The
communication from customers back to the sellers (and other customers) also consists of feedback about
the performance of products. The Sephora app helps get feedback from customers about a product's
performance after days and weeks of the purchase, and it also helps Sephora get information about whether
a particular customer is using right products for her skin type and tone - a factor critical for the success of
a cosmetic product.

11
 Consumers add or subtract
brands as they evaluate what
they want.

The consumer considers Ultimately, the consumer

an initial set of brands, selects a brand at
based on brand perceptions the moment of purchase.
and exosure to recent
touch points.

After purchasng a product or service,

the consumer builds epectations
based on experlence to inform the next
decision joumey.

Figure 3, Evolved loop metaphor customer purchase journey [6]

As sellers have the ability to reach customers by several digital channels and at all time, they can actively
shape the customers' decision journey to deliver value to both the customer and their brand. Companies
can shorten, orin some cases even get rid of the considerationand evaluationphases. (Figure 4). [8]. Some
of the important capabilities that companies need to deliver such value are ability to proactively personalize
the customer interaction and messaging, contextually and timely reaching out to the customers at intended
stages of their decision journey to have best influence on their purchase decisions, and identifying and
delivering new sources of value, such as new services, for both the customer and seller. [8].

12
 START 0,

*TR
HERE

Figure 4, Classic Journey Vs Accelerated Loyalty Journey [8]

The underlying technical capability that companies require is the ability to leverage customer's personal
data and learn to predict their behavior form a continuous stream of data about their activities and presence
on various digital channels. The companies need to leverage machine learning to intelligently automate the
intensive tasks of predicting customer behavior and personalize selling at the level of individual customers.
They can use orchestrate their communication with customers through various stages of the journey, to
improve their sales and marketing spend effectiveness.

2.3 Understanding the customers' purchase decision journey (CDJ) through the lenses of sellers and
customers

The purchase decision journey of the customers largely shapes their experience and perception of a brand,
and so they have the large potential be a source of competitive advantage for improving the sales. This
Thesis adopts from and modifies the traditional CDJ framework. The modified framework (figure 5)
proposes to perceive the decision journey from the standpoint of the sellers and the customers in order to
understand their objectives during various stages of the journey. This understanding will help sellers
strategically align their sales and marketing efforts with the needs and actions of the customers. Following
are various stages of the customer journey that identify various stages of the journey by the goals or
objectives of customers and sellers.

13
 Explore Inform
Be Inspired Trigger

2 2
1
Evaluate the Convince
3 solution
Share Motivate to 6
experience share

&
4 Purchase Loyalty 4 Encourage
5
purchase

Receive Value Delight

Figure 5, Customer Journey Framework: Left - goals of customers, right: goals of the sellers

1. Customers want to be inspired for a purchase;

Sellers want to trigger the desire for purchase

The journey begins typically with a customers' need or inspiration arising from an encounter with a brand,
a product or its advertisement. This stage includes making the customers aware of the brand; the set of
options considered at the outset could be thrice more likely to be purchased finally than the brands outside
that set [6] Today, the customers start their journey with less number of brands in mind due to the exposure
to a plethora of advertisements compared to a decade back. The sellers' objective at this stage is to trigger
that desire for purchase. They want to either make customers feel that their current solution is not sufficient
or enhance the urgency and make them aware of a solution to overcome their problem and solve their needs.
And so, whether or not the customers' consider the sellers' products in their decision journey depends
largely on seller's capability to outreach their potential customers.

2. Customers want to explore options

Sellers want to inform more about their brand/products

At this stage, the customers want to explore options while minimizing their search efforts. The sellers'
objective should be to help them explore options easily by partnering in the research with the customer. A
research by HubSpot about how much part of overall shopping online research is found that half of the
shoppers spend 75% of time conducting online research. The number of options initially considered might
increase as customers proactively seek reviews, ratings, and recommendations, learn about new options
during comparison or update their criteria as a result of learning new information. The recommendations
and reviews have become significantly more important part of the product exploration stage. According to
the Nielson study on the influence of reference on purchasing decisions, it was found that 82% customers
proactively seek referrals before making a purchase. [9]

14
The marketing is increasingly shifting from the time when sellers "pushed" marketed and reached out all
customers with the same content without considering much about the timing and the stage of their decision
journey to the time when customers "pull" information needed to make decisions and inform themselves.
A study shows that the customers engage with only one-third of the touchpoints that are activated by the
company in the evaluation stage of the decision journey compared to two-thirds the number of touchpoints
that are customer driven such as reviews and ratings, social recommendations from peers and influencers
and also include the experiences of past interactions with the brand. [6] In this stage, the sellers focus on
finding on which touchpoints can they can meet their customers and deliver the relevant content through
the relevant touch points. In addition, they should help the customers to move to the next stage at the
decision journey or direct them to the relevant touchpoints for more information.

3. Customers want to evaluate the solution

Sellers want to convince that their solution is superior

While sometimes customers explore and evaluate simultaneously, other times active evaluation of a
solution follows the exploration stage. At this, the customers are clearly intent on purchasing, and their
main objective is to find the best value. Companies' ability to relevantly adapt to different consumer needs
and behavior is based upon their knowledge of customers' profile and this knowledge can significantly
improve customers' experience with seller's brand during their journey. If the customer is new to the
category, the sellers need to inform or educate them about the solutions, type of products, features, brands,
prices and where to buy and where to get the best deal, if the consumer is in the replacement stage, sellers
can update them about their latest products and their features, special value packs, most recommended
products by other consumers or celebrities, competitor brands and prices and where to buy to get the best
deal.

4. Customers want to purchase

Sellers want to encourage the purchase of their product

This stage mainly includes the moment of purchase by the customers. Customers need easy and frictionless
purchase experience. For more expensive or technical products they want the assistance of the sellers during
their purchase process. They also want to be sure that they are getting the best value.

The sellers' objective is to encourage the consumer to close the deal. They also encourage purchase
upgrades, cross-sell, and prompt adding accessories to the purchase to drive basket size up. The purchase
is also called conversion, generally in the context of e-Commerce sales. To improve the sales, the sellers
often drive promotional conversion using tactics including:

15
 - Price discounts: Sellers can drive customers' consideration and conversion by offering price
discounts. However, lowering down the base price in the digital channel can hurt their brand
perception, bring value loss over time and create channel conflict if the discounted price is not
offered in-store or other online channels.

- Coupons or cashbacks: These promotional tactics do not impact base price, and are not picked up

by price comparison engines online. Sellers could use them to drive traffic while reduce channel
conflict and working close to the moment of purchase.

- Bundles: In e-Commerce, not only physical bundles but also virtual bundles can be setup at retailers

without having a cost related to special packaging.

5. Customers want to receive value

Sellers want to delight the customer

After purchasing a product, the customers judge its value throughout its lifetime. Though sellers' objective

post-purchase is to improve the consumer loyalty, it comes as a by-product of reassuring the customers
about the value of the product by helping them to reach 'point-of-value' and delighting them by serving a
best-in-class experience. Sellers could do so by delivering an enhanced post-purchase experience in the

product usage. For example, for the electronic devices they can have a digital channel that provides how to
use guidance, how to maintain/clean/charge instructions, and how to obtain the best results or product life

tips. They need to support the users in case of issues with the products, or the usage, maintenance, repair,

and all the generic customer care related issues. In addition, they should register customers on their database
and track customer activities on social channels in order to growth efficiency their database and carry out
efficient Customer Relationship Management activities to initiate dedicated loyalty programs

6. Customers want to share their experiences and reconsider the brand

Sellers want to motivate to positive recommendations and ensure loyalty

Sellers do inspire repurchases from their customers by creating regular incentives for loyal consumers and
interacting with them post-purchase to communicate the innovative propositions. But because of an array

of options and other sellers trying hard to acquire new customers, sellers need to turn their customers form
'passive' loyalists - who tend to stick with the brand to 'active' loyalists - who go beyond and recommend

it. Sellers should spend on the relevant touchpoints with a focus on expanding their base of active loyalists.
[6]

It is evident from numerous product reviews on the internet that customers often express their opinions
about the products though ratings and review not only on the channels of purchase but also socially. So,

16
after the purchase, the sellers should aim to seek advocacy by stimulating some form of sharing and
promoting using relevant channels. Facilitating rating and review generation will help sellers convince other
consumers in their active evaluation phase by providing them 'social proof of their products, will
contribute to triggering new consumers and also raising the visibility in search engines. Many sellers have
started enabling, stimulating and incentivizing user-generated content about the propositions, so as to fuel
the content on the touch points such as forums, YouTube, e-Commerce platforms, etc. used by other
consumers in their active evaluation phase. It improves the visibility on search-engines and contributes to
triggering new potential customers.

In addition to understanding the goals of the customers at each stage of the purchase journey, it adds higher
value to consider what's happening at these stages more granularly from the lens of the customer. The
following framework (adopted from an article in Harvard Business Review) dives deeper into the process
at the customer end by asking questions about each stage of the journey about four factors

-
Actions: What are the actions of the customers at this stage and what actions do they take to move to the
next stage? [10]

Motivations: Where are the customers coming from? What are their needs? What motivates the customers
to advance to the next stage? What are their main criteria? [10]

Questions: Which questions customers might seek answers to? What do customers surely care for of they
want to move advance to the next stage? Could there be any issues in understanding the value proposition
or technical and operational aspects of the product? [10]

Barriers: Are there potential issues such as a tedious payment process, missing element in the services
with the product, unavailability of inventory, or cost that could hinder consideration of product purchase or
movement of the customer to the next stage. [10]

Following is an example of the application of the above framework for customer TV purchase decision
journey. (figure 6) The seller needs to consider what are all the potential Actions, Motivations, Questions
and Barriers that can come up at customer end during all the stages of that journey. [10]

17
 Awareness Research
- Asks friends for advice
- Visit stores and asks sales person
questions
Activations - Reads websites, blogs, magazines for
reviews and to gain understanding of
* Hears about it from friends or sees it in domain
friend's home e Learns jargons, brands
" Sees it on TV, web or magazine - Look for sales
" House remodeling
" Broken TV
" Super Bowl
Motivations * Wedding gift - Make the best choice
* Children want a TV - Satisfy needs
- Upgrade TV to make it compatible with new * Get the best deal
gadgets to pair with it - Trump neighbors and friends
" Technology change - maybe need a 3D TV - Know enough to not get duped
- Which is best option
* What can I get for my budget?
- Is it future proof?
a Will it work with my other devices?
Questions - Will it be durable?
- What criteria is important to me?
" How much? (Can I afford it?) - Is it user friendly?
" Is it worth it? - Where can I get best offer?
" Is it an improved product? - Will it go on sale soon?
" What will my friends think?
* Honest, unbiased reviews
Barriers - Too much to learn
- No time to do necessary research
- Too many seemingly good or bad
" Status quo - satisfied with TV? choices
" Not tech aware or interested
Figure 6, Detailed TV purchase decision journey analysis from customer lens [10]
To summarize, companies should start with structurally looking into the customer purchase decision
journey and analyzing the objectives from the lenses of both sellers and customers. Next chapter discusses
the touchpoints that customers use to move through their decision journey and process for designing
efficient customer journeys by identifying the relevant touchpoints.
Purchase Out-of-box experience
- Open the package
- Look for directions
- Go to the retailer - Assemble and install TV and
* Compare products peripherals
- Compare warranties * Program and learn the setting
* Consider ad-Ons and - Call customer service
installation services - Share reviews and concerns
online
- Get it done fast
- Check out new features
&Promotion in a store - Relish new purchase
* Release of a new model - Invite friends and trump
- Influencer recommended about the purchase
- Do I need it
* How can I get it home?
- Is it on sale or can I get
discount - Is it damaged?
- What free accessories - Do I need help and directions
accompany the TV? to set up?
* Is TV compatible with my - What do I do with the old TV?
older accessories? - How do I use it and change
- Is it in stock the settings?
- Does retailer deliver it? * Is it meeting my expectations?
* What to do with old
- Don't have desired model peripherals and furniture?
in stock - Not compatible or too
- Unhelpful/unconvincing complex for pairing up with
sales person other devices
* Discovery of bad reviews e Involved set up time
- Offer on competing brand consuming and no help from
customer care
18
 3. Customer touchpoints

3.1 What are touchpoints?

Touchpoint Metrics (2003) defines touchpoint as "every point of contact-online and off; each
communication, human resource, branding, marketing and salesprocess initiative createstouchpoints. The
quality of touchpoint experiences drives perceptions, actions, and relationships." [ 11]

This definition of touchpoint provides a brand level definition of touchpoints. Accordingly, a touchpoint is
every point of interaction between the company and all its stakeholders outside including the customers.
One can conceptualize the brand level touchpoints using Denise Lee Yohn's framework called 'Brand
Touchpoints Wheel' (figure 7, presented in her book 'What Great Brands Do') for various teams in a
company to systematically identify and assemble their brand's touchpoints. [12]

-4

~Og V

US /~/ /

Figure 7, Brand Touchpoint Wheel for all company touchpoints

Teradata (2001) defines touchpoint as "customer interaction channels such as call centers, web sites,
automatedteller machines and web kiosks." [ 11]

For the purpose of understanding the touchpoints in the customer purchase decision journey, we will go
with the Teradata's definition of touchpoints. So, customer touchpoints in our discussion will be considered
as any point of interaction between the company selling products or services and the customers during their

19
 purchase journey. In 'Brand Touchpoints Wheel' framework that was originally conceptualized by Davis
and Dunn (Davis and Dunn, 2002, p 6-7), customer touchpoints are categorized and explained in three
stages as follows: "the pre-purchase experience touchpoints, the purchase or usage experience touchpoints
and the post-purchase experience touchpoints". [13]

Brand Touchpoint Wheel

Post Pre-
PucaePurhase Adverbsing
Expenernce Expenence

Newsletters Purchase P

'
Experienrce

Sales Force Pouuevc

Assorment

Figure 8, Brand Touchpoint Wheel for customer touchpoints

3.2 Mapping touchpoints

This thesis explains the Customer Purchase Decision Journey in six stages. By adopting the idea of mapping
the customer touchpoints to various phases of their purchase (Brand Touchpoint Wheel, Davis and Dunn),
map all the touchpoints to various stages of purchase decision journey. This derived framework considers
2X2 matrix to map touchpoints across the stages of the journey, because it makes the possible mapping of
multiple touchpoints with multiples stages of the journey. In reality, the customer could be interacting with
same touchpoints in different stages of their purchase decision journey. In addition to mapping the
touchpoints using this framework, the companies can map the existing and potential customer journeys
through those touchpoints to ensure they move through the stages to purchase and share as intended by the
companies. Following table maps potential journeys of a furniture buyer:

Customer's Be Evaluate Receive

Objective Inspired Explore the Purchase Value Share and Reconsider
solution
Seller's
Oeie Trigger Inform Convince Encourage Delight Motivate to Share and
Objective Purchase encourage Loyalty

20
 Google
Search \

Online Ad -X
Social XX
Native Ad
e-commerce X
Website
Reviews 4 X

Lifestyle
Forum
Direct email
Brand
Website

Figure 9, Customer journey maps through touchpoints

Mapping would help sellers evaluate the journey by understanding the following: [14]

- The points where customers face obstacles

- The message which is difficult to understand or the information that is complicated for customers
- The points in the journey where consumers lose engagement or interest in the brand or the product
and would prefer to not engage with this brand
- The points that provide consumers most satisfaction to motivate them to move to the next steps

3.3 Designing the journey

Customers' experiences about a brand are created when they interact with its touchpoints, and that
experience plays an important role in whether they would consider that brand for their ultimate purchase
decision. The goal of designing the customer journey is to design the customer experience a brand intends
to provide to its customers. Today, it is not enough to only be reactive to consumers during their purchase
decision journey but be proactive to influence their decisions by designing it. [8] Good journeys have
potential to provide a competitive advantage to the sellers. Companies who optimize theirjourneys can not
only acquire and retain more customers but also provide value for themselves and the customers.

21
Bellos and Kavadias (2011) state that "the optimal design decisions depend non-monotonically on
two important parameters:
i) the variability of the experiential outcome from touchpoint interactions and ii) the underlying
interdependencies between tasks, which may give rise to correlated experiential outcomes across
touchpoints." [13] Customer experience is partly a function of variability caused by multiple ways a
customer can move across the touchpoints and different quality of experience each touch point provides to
a customer. It is desirable for the sellers to present a consistent customer experience by optimally designing
the journey across touchpoints in order to positively determine the journey outcome. A well-designed
journey will consistently, effectively and clearly communicate the brand across all the touchpoints
throughout the journey.

3.4 Considerations for designing the customer journey

1) Identify key touchpoints

For a particular product, a seller needs to identify what are the primary touchpoints that its existing and
potential customers use during a particular stage of their purchase decision journey. Sellers can choose for
several touchpoints these days. Some of the following criteria will help select the touchpoints over others.
The effectiveness of the touchpoint, which depends on the reach of the touchpoints among the customer
segments that seller intends to target, and the size of the segments interacting with those touchpoints. The
touchpoints also need to be assessed by their usage on various devices if the seller wants to interface with
the customer on particular type of device - the customers are more likely to use and get influenced by
different touchpoints on mobile, desktop and tablets. For example, a certain type of content and touchpoints
are more suitable for mobiles such as the Ads and in-App notifications triggered based on geolocation.

In most stages, there are multiple touchpoints that the customers use to achieve their goals. Also, many of
the same touchpoints could be used in more than one stages of the journey. How can a seller prioritize the
touchpoints to be used? Attribution of sales to various touchpoints using data analytics can provide insights
into the contribution of each touchpoint. Tools such as Google Analytics Solutions use data-driven machine
learning approach to determine the impact of each touchpoint in the customer journey. [15] However, the
caveat is that the insights from attribution should be derived not completely on the face value of the
attribution percentage that the tool provides, because the impact of many non-digital touchpoints
responsible for influencing customers' purchase decisions such as billboards cannot be measured. Also, the
impact of some of the digital touchpoints such as Display Advertisements and social media Advertisements,
which might create repeated visual impact but is not clicked by the customer might not be attributable but
does have a role in the purchase. This touchpoint might indirectly create impact; however, some other

22
touchpoint might get higher attribution despite its lesser role in creating the influence on customer's
purchase decision.

2) Identify the role of the touchpoint in each stage of the Customer Purchase Journey

The questions to ask about the role of touchpoints with respect to the customer journey stages are:

- What specific things are sellers doing at each touchpoint?

- Is a particular touchpoint answering customers' questions and addressing concerns? Are they
targeted to desired customer segments?
- Are the touchpoints going beyond addressing the needs and delighting the customer's unstated
needs that the competitors are not? [16] The role of social touchpoint such as Facebook
Advertisements is to trigger the need for purchase and spread brand awareness in the exploration
phase, direct the online customers looking for discounts and deals to the sellers' website to motivate
them in their exploration phase and provide social proof of the likeability of the brand and
recommendation (measured by the Likes and Shares) in the evaluation phase and post-purchase
loyalty phase. Identifying why customers are using particular touchpoints during a particular stage
of the purchase journey helps sellers prioritize the touchpoints based on how efficiently they meet
the desired objectives of the sellers and goals of the customers. It also helps identify the content
needs of the touchpoints ad adapt the message/creative to those touchpoints.

3) Identify what should be the next action of the customer

By predicting the next move of the customer using data and the stage of the purchase decision journey the
customer is in, the seller can guide the customer to the next relevant touchpoints. The seller can also identify
if there are specific hindrances in the journey from one stage to the next that led to disengagement of the
customers' or dissatisfaction that calls for use of more expensive steps to retain or attract the customer back.
[16]

For example, if data about a particular customer shows that she is evaluating options for purchasing a coffee
machine, the seller could guide her to the review video of the coffee machine of its brand, or to a
comparative chart of various coffee machines by variants of its brand or all comparable brands. These
touchpoints could in-turn guiding them further to the online or offline channels for purchase, creating a
seamless journey unlike the broken journeys we see online today.

Example: YouTube (people seeing a music video) 4 banner ad (with clear call to action) -- clicking leads
them to Retailers YouTube channel (where seller describes in detail the product's value proposition)

23
For motivating the customer's potential next steps through the touchpoints, the seller needs to determine:

- Which stage of the journey customers are in, what is the relevant next touchpoint that the seller
wants them to take based on sellers pre-designed customer journey for that product, and what is the
'call to action' that a maximum number of the target customers would respond to.

So, by deciding upfront how various potential journeys might look like for their customers, sellers can lead
the target customers to the specific content and touchpoints that motivate the customers to progress into the
next stages of their journeys.

Following are some of the pitfalls that the journey designers need to consider: Though the prioritization of
touchpoints is important, planning such that the customers who do not land on those top predicted
touchpoints do not make into the end of the journey would be a rigid journey design. So, having a minimum
viable ecosystem of touchpoints in place is important for pulling into the journey maximum set of target
customers. Also, in order to shorten the customer journeys to bring the customers to purchase phase faster,
the designers could make the mistake of cutting short the journeys at wrong stages, losing the customers at
those stages. So, the call to actions on all the touchpoints should be considered as bridges in the customer
journeys and designers should check for any shortcuts in their designs.

4) Incorporate the understanding of various customer personas

"Personas are the starting point," says Michael Hinshaw, customer experience strategist and president
of MCorpCX. "Because a journey map is the story of a customer" experience. [17] Customer personas help
develop a user-centric design of journeys and help think about the possible variations of customer journeys
comprehensively. As real research and customer interviews are conducted to identify personas, they provide
an additional basis to identify various possible touchpoints particular customer persona might be more
inclined to use while interacting with the company over the other touchpoints which company might
originally expect of the customers.

To understand how the touchpoints and the customer journey differs for different personas, refer to the
following example in which Persona 1 purchases phone for personal use as the desire for new technology
triggers her compared to Persona 2 who purchases phones for her corporate employees for their (primarily)
office and personal use. Though example uses a high level understanding of the two personas, sellers should
look at the potential customer journeys of all major personas more granularly.

High-level Persona 1 description (figure 10) - purchases for personal use, digital savvy, frequently
purchase new phones, performs research and comparison on online channels as well as prefers to have the
touch-and-feel of the phone in-store (Omni-channel research)

24
 Desire for Google
Newest Decides to h
Price Ii
Technology etail Store Research Decides to Buy Comparison Samsng
Wvebsitessc
Samsung'so

Am Tech
Amazon Retail Forums
In-store Digital ad location
associate
YouTube Customer
Service

Figure 10, Purchase Decision Journey of Persona 1

High-level Persona 2 description (figure 11) - purchases for corporate employees, does moderate online
research but depends mainly on the agents and dealers for a corporate deal, uses the corporate channel of
purchase, bulk purchases new phones, may require to replace and repurchase phones in future.

Decides to
Research
Need to purchase Samsung's Google
phones for office Decides to buy Corporate Order
web
employees search placement channel

Employee an
R*epurchase
and
phone.
Agent
feplacemequ
pe i6stributor Distributod' Agent
Distributor Location
Retail sites
location
Agent Distributor

Figure 11, Purchase Decision Journey of Persona 2

3. 5 Touchpoints - an integrated view

A media company was facing the problem of retaining its customers and acquiring new ones. The company
executives decided to improve the customer experience to improve the situation. But when they measured
the customer experience, they found that score for each of their touchpoint was strong 90 percent. They
concluded that their service was great. However, a focus group in their company contrarily reported that
the customers left their company because of poor customer service. Upon deeper analysis, the executives
understood that the overall satisfaction that was a result of cumulative experiences of customers across
touchpoints throughout their journey with that company was low, although each touchpoint individually
performed well. If a customer interacted with four touchpoints through their journey, each with a 90 percent
chance of going well, the overall customer experience in their journey fell by average 40 percent. So though

25
 touchpoints performed well, it did not ensure great end to end performance. A supporting evidence was
found by McKinsey & Co. in one of their surveys - "the gap on customer satisfactionbetween the top- and
bottom-quartile companies on journey performance was 50 percent wider than the gap between the top-
and bottom-quartile companies on touchpoint performance." [18] It is evident that more than the
performance of touchpoints, the performance of journeys is much more effective at driving business
outcomes such as churn and repeat purchases, and thus revenue. It is because journey performance is
strongly correlated with the total customer satisfaction. The following exhibit shows results of a survey that
explain the correlation between customer satisfaction and their willingness to recommend the brand with
touchpoints Vs journeys. (figure 12) [18]

Journeys are significantly more strongly correlated with overall

outeornes than are touohpoints.
0 Vskms'
*1buchvoftz *Jaa,10rot

Customrer asalmon W jecommind

tngnem

U .4
0.52
0,23 04
0.47
P1 wa +17% 0 +104%

Hoftblm030 0.28
0-53 049

HotJso3y [ 8

+56% +1%

00s4*vts .r, 0 vwoft o a'mUi~P*-M9%S moo rs~4j~s~na 1 c

Sower, Mrn.y k 1 cms-ivduks' rnc voy. Juw.-Od 2015 data

Figure 12, Correlation of customer satisfaction and Willingness to recommend with touchpoints. Vs
Journeys [18]

3.6 Important touchpoints for Ecommerce

3.6.1 Online Advertisements

- Search (organic)

Most consumers use search engines to find what they need. Search engine (e.g. Google, Baidu, Yahoo)
results are based on content available in other touchpoints after a consumer made a search query. The results

26
of this search are based on the rankings of search engine algorithm. Organic search is the critical touchpoint
for every Customer Decision Journey (CDJ).

Presence in the Search Engine Result Page (SERP) of the results is a powerful way to trigger customers
and to create brand awareness and initial consideration. Sellers can influence these results by doing search
behavior analysis. It needs the right content in the right places of the CDJ to work efficiency. So,
accordingly produce relevant content and ensure that they deployed it in the key touchpoint that has been
identified. The keywords and phrases that customers use while searching particular information about a
product category or brand need to be identified from sources such as Google trends, etc. because search
engine displays the results of the search query based on keyword matching.

- Search (Paid)

When users put search query in the search engine such as Google, the initial few search results associated
with that search are sponsored and indicated as ads. These web pages pay to appear as top results of the
search query. [19] Paid search works very well in the trigger phase. If sellers do not have the content in
the right places, the invested money in paid search will be wasted (not efficient) because it will drive people
to a touchpoint that does not have relevant content and customers will not convert.

- Display

Display advertisements are the form of advertisements that appear on websites or search engine results
pages such as publishers (example New York Times) or social media (example Facebook). [20] Display
advertisements are also called Banner Ads. Various forms of Banner Ads are Traditional Banner Units,
Mobile Banner Units and In-Banner Video Units. These ads are scalable and easy to deploy, and results
can be measured to a high degree. Banner Ad campaigns are measured using the following metrics: Clicks,
impressions, conversions, spend and video views.

- Video

Video advertisements are run either before, during or after the online video content. It enables advertisers
to connect their message to an engaged audience. Also, as the draw of viewers towards internet TV is
increasing, and the video ads are spanning across four screens - computers, smartphones, tablets and
connected TVs, the potential of these type of ads to outperform other types of online ads for branding is
increasing. Watching the video online has gone from a niche activity to mainstream and Video click rates
are far higher than image format ads. Users click on video ads about five times as often as they do image
ads. [21] Major performance metrics for video ads are 100/75/50/25 % Completed Rate, Impressions,
Clicks, and Click-to-rate (CTR).

27
 Mobile

Mobile advertising involves displaying text, graphic images and animated ads on data-enabled mobile
devices. These have transitioned from being a small format of the web display ads to being a separate
category of ads. Mobile advertising is becoming a more crucial part of the marketing mix for brands of all
sizes, as companies develop more sophisticated methods to engage consumers, influence their buying
behavior, and even complete sales via a mobile device.

The major types of mobile advertisement as are: Click-to-download ads - the user is directed to the Appstore
or Google Play, Click-to-call ads - the user calls to a phone number after clicking the button, image text
and banner ads - a click opens users' browser and re-directs them to a page, and push notification - messages
that pop up looking like updates in the notification area of the devices. They are generally pushed from the
Apps that are installed on the devices. [22]

The major types of mobiles ad formats for Mobile Web are Optimized - in which the advertisement is
responsive to the device size. In this, the format of the Ads reformats itself to suit the screen dimensions.
Non-Optimized - some websites still have their site layout independent of the screen size of the mobile and
the advertisements are called. Another common type is In-App mobile ads that could include the banners
and expandable ads. The metrics used for measuring the performance of these types of ads are same as
those for display ads.
Some of the newer formats of mobile ads are as follows:
- Interstitials ads that appear intermittently during the app usage and are larger than other ads to allow
for more engagement of the mobile users. [23]
- Native mobile ads that coherence in formatting and content with the hosting app or site.
- Deep Linking that are actually the URLs of the mobile apps. These ads have an embedded link called
Uniform Resource Identifier (URI) URI which an address for an app or a location within the app, and
is used from bringing the potential customer to a particular location in an app such as a product page
of a particular item or a description page of a deal. [23]
- Beacons are external devices that connect to the app using device's Bluetooth technology. They detect
nearby smart devices and send them media such as ads, coupons, location-based deals, or additional
product information.
- Native
Generally speaking, native advertising is advertising that fits 'form and function' into the context of the
webpages on which it appears. For example, the advertisements that appear as the sponsored feed or news

28
on Facebook along with the content posted by the regular user and Facebook friends are the Facebook
native ads. In one of its reports, the Interactive Advertising Bureau (IAB), the primary organization
responsible for developing ad industry standards detailed six categories of native ads: [24]

- In-Feed Ad Units: These units appear as normal content feeds of the website, as if the content is written
to match the surrounding stories of the website.
- Search Ads: Paid search ads are also considered native ads because of their nature that they appear as
the part of the main content of search engine, but are generally above the organic search results.
- Recommendation Widgets: These form of native ads are also part of the content like In-Feed Ads,
however, they are differentiated using external references or indicators such as "Elsewhere from
around the web", "You may have missed", etc. like suggestions.
- Promoted Listings: These are similar to the paid ads in that they appear like the main content of the
website on which they are hosted but are sponsored. For example, on the food delivery ordering website
such as Yelp, some of the restaurants appear on the top of the search results along with other search
results below them.
- In-Ad (IAB Standard): This ad fits IAB standard size container found outside the feed, containing
"...contextually relevant content within the ad, links to an offsite page, has been sold with a guaranteed
placement, and is measured on brand metrics such as interaction and brand lift."

- Social Media

Advertisements on the social media intent to display content that the social media consumers would interact
with (share, comment, like, retweet, play, click and take subsequent action) on their social networks. The
channel partners gather the data about all these interactions to evaluate high-value customers for a particular
type of ad and target them and other similar prospects. The key players in this category of ad publishers are
Facebook, Instagram, Snapchat, Twitter, Pinterest and Linkedln.

- Blogs and News (Inbound Marketing)

Sellers reach the potential customers who might have high interest in a particular category of product
through the blogs, articles, news or reports about these categories or topics related to these categories. For
example, if a customers need a new smartwatch, they generally want to compare the brands or read reviews
on blogs by technology product experts. These articles could include recommendations, features and finally
links to websites of smartwatches that sellers want to attract customers too. Sellers often partner with blog
writers and reviews who have high number of followers to get them to write sponsored articles. E-commerce

29
websites such as Amazon often drive customer by sponsoring articles such that might be about '5 coolest
products on Amazon you don't know about'.

HubSpot, one of the leading companies that provide inbound marketing tool and training to the sellers, has
defined a four-step methodology for inbound marketing. First, attract right people, ones who are more
probable of becoming customers. This could be done by writing educational blogs that answer the potential
questions of the customers, placing this type of content on the places where people are more likely to search
it, and placing high quality and valuable content on social media to get it shared. After people visit the
website through the links, the second step is to convert them into leads by personalizing the engagement
from there on by understanding their requirements through forms, chatting or exchanging messages with
them, providing them easy and quick ways to reach out to business for any queries and systematically
documenting the interaction with them using the Customer Resource Management tools. Once the business
has leads, the third step is to convert them into customers. In this step, identifying the best leads and
nurturing them is important. Not all leads are qualified for the investment required to convert them, so
predictively scoring the more qualified lead is recommended. Further, understanding the stage of purchase
they are in and communicating with them by providing focused and relevant information. Finally, in the
fourth step, the framework recommends staying connected with the customers after converting them by
sending out the content smartly based on their product lifecycle stage and customizing the conversations
by keeping track of comments and requests. [25]

ATTRACT CONVERT CLOSE DEUGHT

Figure 13, Method for inbound marketing [25]

3.6.2 E-commerce website

E-Commerce sites encompass a full CDJ, hence they have a role in all the phases of the purchase decision
journey. Many people start their journeys from the search page of the e-commerce website or the catalogue,
and end it with the purchase on that website and become repeat purchasers and sometimes loyal members,
which is the case with Amazon Prime members. Various touchpoints on the website correspond to stages
of the CDJ such as: Search page -> Exploration stage, Product Recommendations -> Trigger, Product Page
-> Inform

The sites themselves are massive touchpoint that the customers visit, therefore e-Commerce sites have wide
reach and can serve as media platforms to buy media and do advertising. E-commerce sites span across

30
different ecosystems such as mobile, desktop, r-reader (for example, Amazon.com, Kindle, Amazon Mobile
app), and so they attract huge consumer base and are usually in the top 10 most visited sites of the respective
country. Hence the media deployed in these entities get a broader reach. In addition to the reach, as
customers are in the mindset for shopping, the CDJ is shortened and probability for call to action is stronger
to move the shopper towards the purchase and next stages. Finally, e-commerce sites profile their
consumers and their shopper behavior, that enables targeted and customized communication possibilities
(e.g. Amazon Media Group deploys the media banners to different shopper personas based on lifestyle such
as vegan, young professional etc.).

Some of the strategies that e-Commerce websites employ following strategies to increase their conversion:

a) For banning:
- Seamless execution of banners across platforms exposing customers to multiple and consistent
exposures. For example, a customer who has seen Nike ad on amazon.com sees the same ad on Amazon
mobile website or kindle
- Rich banners on static ads: For example, ratings and reviews are embedded in banners, directly add
to cart or wish list button are embedded in banners, use of expandable ads that extend the banner when
hovered over that shows complete product detail or category pages are enriched with how-to videos
and other relevant content
- Focus on KPIs and objectives are measured that are linked to business strategy (for example, cost per
action)

b) Next to the bannering there are other ways to trigger consumers within e-Commerce sites:
- Internal search: product-related search is usually higher in volume than the amount of search in search
engines.

- Deal of the day, Top sellers

- People who bought also bought, products recently viewed

- Bundle suggestions. Example, accessories such as sleeve, headphones, wireless charger, etc. with tablets

- Buying/gift guides

- Display network

Search engine display network: Next to paid and organic search, search engines also own display
networks. For example, the basket of placement in websites or high traffic portals that the ads are deployed
to through rich or static display banners. Display networks of search engines work almost exactly as display
networks of media agencies. One key advantage of display networks that are run by search engines is their

31
power of retargeting. Search engines, based on the queries of their visitors, can show the display banners
on the websites that are relevant to the shopper at the right place at the right time.

E-Commerce Site Display networks: In addition to displaying ads on their owned web entities, e-
Commerce sites also own display networks. An example is Amazon media group display network.

To summarize, customers interact with several touchpoints in the purchase decision journeys. The
experience of the customers and the success of sellers depend on how well the journeys are designed.
Further, it is important for companies to know various touchpoints on which they can meet their customers
and should map the potential journeys based on which companies should plan their sales and marketing
efforts. Next, the thesis discussed the technology and working of the above category of touchpoints in the
next chapter.

32
 4. Technologies Underlying Touchpoints

4.1 Recommendation Systems

Recommendation Systems use algorithms that predict the customer preference for any item on the website
and recommend those items to the customers. [26] They are meant to be profitable for the customers as
they reduce search cost and time, minimize risk while maximizing profits. [27]

Recommendation systems help improve the sales on an e-Commerce website in several ways throughout
various steps of the customer journey. Most importantly, it has the potential to convert a visitor into a
customer. When a customer is in the product discovery phase and lands on the e-Commerce website, the
landing page itself, when personalized to display and recommend relevant items, drastically improves
customer experience and reduces effort of searching the desired product. During the evaluation phase of
the journey, when customer understands the product features in more detail (generally on the Product Detail
Page), the recommender system can increase sales by cross selling and up selling by bundling the product
with accessories or peripherals which are based on the information about past purchases of customers who
bought multiple related items in a single purchase. It also improves customer loyalty due to the value
addition made by websites that they can do by learning about the customers and providing personalized
recommendations and communications consisting of recommended items.

Larger companies are heavily investing to improve the personalization of their products. Recommendation
systems are a big part of that effort. For example, every year Netflix makes $150 million investment to
improve their content recommendations as it has very small amount of time to convince a user to watch a
particular piece. Larger e-Commerce companies make recommendation engines in-house. However, may
e-Commerce players prefer plug and play solutions slightly modified for their products, whereas smaller
players use recommendation software as a service. [58] Some examples of recommendation products are
Adobe Target, Recommendations API offered by Microsoft. Companies, and recommendation engines
offered by Barilliance and Strands. [59]

4.2 Applications of recommendation systems in e-Commerce website

Some examples of the applications of recommendation systems on the e-Commerce website such as
Amazon.com and eBay are:

Different types of recommendations on Amazon.com (Figure 14 and Figure 15):

33
Related to items vou've viewed see more
Inspired by your shopping trends

Figure 14, Recommendations on the landing page of the e-Commerce website [28]

Inspired by your browsing history see more Frequently bought together

- Tota price: $283.26

- 2 0

This Item: Samsung Galany J7 Pro (32GB) J730G/05 (Black ) S.S* Ful 140 Dual SIM Unlodked Phone with... $266.59

5 13-PACKI- Mt Shield For Samsurig "Gataxy J7 Prn" (Wilt NOT lc Lfo( 17 Sky Pro' (Tempered Glasa.. $7.99
Galaxy 17 Pro J730G Case.Marna Mouth Shockroof Hedvy Oury Combo Hybhid Rogged Dal Layer Grip Cower .. S8.88

Customers who viewed this item also viewed Compare to similar items

Samsung Galaxy J7 Pro Samrsunria ,. j, Prme

SM-J730GM/DS 32GB (32G) G61DFDS - SS'
This item Samsunq Galaxy J7 Pro Samuing Calaxy J7 Prime Factory
Black. Dual Sim, S.5" Dual SIM UnLocked Phone
(32GB) J730G/DS Black) .5' Full Unlocked Pone Dual Sim - 16G0

-
13MP, GSM Unlocked... with Finger Print Sensor... HD Dual SIM Unlocked Phone with White Gold
700 FingenPrint Sensor (US anid Latin
$269.88 cprkne $259.00 LTE)

Figure 15, Recommendations on the Product Detail Page [28]

Figure 16 shows a feedback recommender system on eBay.com. Buyers and sellers provide feedback on
this system after doing business with each other. The feedback includes rating and comments. This
recommendation can be pulled up by either party to verify the other before engaging.

34
 Hi. Jdebord!{SIgn A) Se Map

Cat"godee Morse Expreta Stores

Home > Commumy > Feedbeck Formw : Feadhack Proe

Feedback Profile S*,

Positive PFeedck (lest 12 monttw): .4% PHow is Foedback Perranteg. caleted?]

Mvwsira Jt-14-04 It Untod Slates

--------- S.mber .ng. uMemberUQukd Unks
Recent Peedbacit Ratrag eald e ags t ma Mombar Quick ths
1 2Mem"e hutvnn coi65t #mt

ltem adescrib.d *****A 525 View m porms.

ar Feedbackpmo f, rs on

4.3 Approaches to recommendation systems

Recommendation systems are one of the subclasses of the information filtering system. They are filtering
problems because recommending an item to the customer is like discarding the not useful or unimportant
items. [30] Typically recommendations have been produced in two ways -- Collaborative and Content-based
filtering. Often the above-mentioned filtering approaches are combined to form Hybrid recommender

systems.

4.3.1 Collaborative filtering: This mechanism of filtering recommends items to customers based on the
preferences of similar customers. The algorithms of these systems collect behavior data of the
customers, their activities, and their preferences, ratings and/or remarks to predict what similar
customers might like. [31] The key idea is to find the notion of similarity between customers of an
item.

The main advantage of these systems is that they do not depend on the understanding and analysis
of the products or services it considers for recommending. The disadvantages are that these systems
depend on the large extent of customer data to make useful recommendations, and so
recommendation systems of e-Commerce websites that use collaborative filtering approach might
suffer from 'cold start' due to the lack of data in the beginning when there are fewer customers.
Another disadvantage of these systems is that only a few out of several thousand products on the

35
 e-Commerce website are liked or rated by the customers, also termed as 'sparsity', so the number
of products that the system can recommend to other customers is limited and most of the products
remain out of or at the bottom of the recommendation list.

Collaborative recommendation systems are of two main categories: memory based and model-
based.
- Memory-Based: It is called memory-based approach because it consists of storing all the customers'
information and predicting the rating or preference of the target customer by retrieving the information
of customers similar to the target customer. This approach has three steps. In the first step, the
algorithm evaluates the similarity between the target customer and other customers with respect to the
item under consideration. Next, using the similarity as weight, the algorithm finds the nearest neighbors
(similar customers). Finally, it predicts the preference of the target customer using similarity weights.
(Figure 17)

Evaluating similarity between target user and

training users

Target user-centered formation of nearest

neighborhoods

Score prediction using similarity of nearest

neighborhoods

Figure 17, Steps for memory-based approach [32]

The weighted mean is one of the algorithms generally used in the memory based approach.

vaj = k w(a,)vw k= / w(ai)

In the above equation, a is the customer to whom the item is to be recommended any object o

The equation finds a rating to predict how likely customer a is to like object o. It basically combines
normalized rating of users. Here the sum is taken over all users. As not all users contribute equally to the
average, the weight controls the inference of all customers on the prediction. This weight is related to the
similarity between u, and ui. The more similar both ua and ui the more the contribution user ui will make in
predicting the preference of user ua. Here, k is a step for normalization that takes the sum of weights of all
users.

36
To determine the weight function w, there are many approaches. One of the widely used approaches is
Pearson Correlation Coefficient which measures the linear relationship between two variables. For
example, whether two customers both tend to give lower or higher rating to a similar item. [32]

- Model-based: Memory-based approach is generally real time and so not very fast or scalable because of

the extensive datasets of continuously generated customer data. A model-based approach attempts to
overcome these problems. In model-based approach, models are build based on the partial data set instead
of the complete data set for making recommendation each time. [33]

There are a number of techniques to build models. Some techniques are:

* Probability problems: It is a probability problem when a task of recommending an item is carried

out by predicting the probability of a preference rating taking a particular value. Clustering models
such as K-means and Bayesian network are some examples of probability techniques for the model-
based filtering.
0 Enhanced memory-based technique: This is the enhancement of the concept discussed above in

the memory based approach which predicts the customer's preference rating for an object by using

the similarity between customers as weights. The idea is to build the model based on a particularly
limited number of data points enough to provide the desired accuracy and store the similarity

measure to predict customer rating.

4.3.2 Content-based filtering: This approach is based on the item similarity rather than customer

similarity. In this system, the products or items are identified by keywords. Also, the system makes
customer profiles by considering the products that customers like and then recommends products

that are similar to the ones customers like. The basic process followed in this approach is of
recommending the best-matched products with the ones rated in the past.

The architecture of the content based filtering system would look as shown in Figure 18. The item's

information is the input to the system (read Doc Source) which is classified by the binary classifier.
The binary classifier has data about the interests of the customers which the system keeps track of.
There is a utility function that works as a decision module. It evaluates which items will pass

through the binary classifier based on its utility based decision-making criteria. This utility function

sets a threshold value that the items need to measure over to be recommended to the customers. It

is not based on the ranking of items but measures utility absolutely. The initiator module takes the
input such as customer's keywords, etc. which are input along with the attributes of the initiator.

37
 Initialization +

"
Accepted Docs
...
um.~~~ Classifier iiii JUe
User
Doc Source
utility func

Figure 18, Architecture of content-based system [34]

Content-based recommendation approach would do reasonably well in practical applications and
could be a starting point because it is dependent on a limited number of customer ratings. It can
recommend items based on its similarity to other items even though it has not been rated, so there
is less dependency on all the items on customer ratings.

One of the most used techniques for content-based filtering is similar to TF-IDF (Term Frequency
- Inverse Document Frequency). The algorithm abstracts the text-based attributes of the products
in the database of the e-Commerce website. The product is weighed based on the frequency in
which the customer's searched keyword for product search matches the product attributes in the
database. [35] However, nowadays sophisticated Machine Learning techniques including decision
trees, clustering, Bayes classification, etc. are taking over to predict the likelihood of customer
liking a product. [26]

4.3.3 Hybrid recommender system: Hybrid approach generally combines collaborative filtering
approach with content-based filtering. It might also unify demographic and knowledge-based
approaches in the hybrid system's model. The demographic approach includes demographic
attributes of the customers, and this information could be very useful in predicting customers'
product preferences when they vary more by locations. The knowledge-based approach is a more
advanced one as it infers the customers' need and might use the functional knowledge of the
products that is abstractable by the algorithm to recommend better products that meet those inferred
customer needs. [26]

Several approaches are used for hybridizing the techniques such as:
- The weighted approach in which the preference score of a recommended product is a weighted sum of

the preference scores of various products. The fundamental logic behind this approach is to optimize
the weights for set of products using the past data [31]

38
- Switching approach in which the recommendation system switch between results of different
techniques and chooses one to recommend. If same product recommendation is resulted by using two
different techniques, say both content-based and collaborative filtering, then those products are ranked
higher
- Mixed recommendation approach in which the top-ranked results generated from various approaches
are presented together
- Meta-level approach in which one of the techniques first generates a model that becomes an input for
the next technique

The studies have empirically shown that the unification of approaches in the hybrid system produces more
precise product recommendations compared to individual approaches, and also overcomes problems in the
individual approaches discussed before such as 'cold start' and 'sparsity'. The problems with the hybrid
system are that it relies on the quality of integration of various approaches such that the limitations of
combined approaches are not inherited. Another issue is to update the recommendation system by taking
updated inputs of customer rating for the products. The key to the problem of making a model adaptable is
to use Machine Learning algorithms to incrementally train the model using new inputs. [31]

4.4 E-Commerce Product Search

Numerous purchase journeys start with the product search when the customer inputs a query in the search
bar of the e-Commerce website. Thus, optimizing the search results that are the result of product ranking
algorithm saves customers time and effort of discovering the desired product from several hundred
products on the e-Commerce website. Ranking algorithm scores the products in the database and presented
to the customers in the order most relevant to the user. Ranking algorithm factors many features for ranking
the products. The types of features are: search query features - the product description keywords in the
search text entered by the customer, user behavior or interaction features such as revenue from a particular
customer, clickthroughs, number of add-to-carts, etc., product features such as price, brand, variations,
ratings, and reviews.

Product search is an information retrieval process in which the objective function is to retrieve the product
in an order that maximizes conversion (product purchase). For a particular search query, the objective
function combines the various features such as product, query and user behavior features with proper values
of weights assigned to them to construct an optimized ranking of the products. That is, the probability that
a particular product is relevant to the customer is a function of the above-mentioned features. Assigning
correct weights to various features in the function is a challenging task where Machine Learning algorithms
can contribute.

39
Following features (not an exhaustive list) might be used as inputs for the search algorithm for e-Commerce
websites [37] :

Product features:

Price: Algorithm makes use of the price to score a product

Title: Product titles are matched with the keywords used by the customers during the search

Image: Having a high definition image of the product that could be zoomed counts in the algorithm.
However, animated images that show 360-degree view or how to use videos may still not be the part of the
algorithm for search results

Variations: The variants of the same products such as sizes or colors have more importance in the algorithm
because multiple customers might look out for variations of the product which they searched

Customer Behavior features:

Type of membership: For example, on Amazon.com amazon prime members (subscribers of their loyalty
program) might have a different result compared to the nonprime members. The prime members might be
displayed items that are eligible for prime delivery at a higher rank

Purchase history: Various parameters of past purchase might form key inputs about customers behavior

Key business metrics:

CTR (Click through rate): This feature could be more important as a feedback feature. If some products
are more interacted with by clicking and proceeding through stages of the customer journey, they would
be considered more important by the algorithm.

Sales Rank: If a product has sold more, it has higher probability of selling more

Learning to Rank (LETOR) method has been used as a widely used method not only for web search but
also for e-Commerce search.

Other features:

Stock-outs: If products are stocked out, they will be ranked below in the search results

Seller rating: Products sold by poorly rated sellers will be ranked low

The above features are the input to the algorithms of Machine Learning model that is discussed in the next
chapter.

40
 4.5 Advertising

4.5.1 Programmatic Advertising

The advertising ecosystem has changed from traditional manual transactions of advertisements between
advertisers and publishers to programmatic. Traditionally, buying and selling were made by salespeople
who made Request for Proposals (RFPs) and negotiated prices. Now, the programmatic technology has
enabled automatic and thus efficient tractions in the space of digital advertising. Before getting into the
process, it is important to understand the main players, components and key technologies in the ecosystem.
[38]

AdSretws wA efSied

Buy Side Sell Side

Figure 19, Representation of Programmatic Advertising Ecosystem [38]

4.5.2 Key players and components of the advertising ecosystem

The advertising ecosystem (Figure 19) has two sides - buy and sell side. The buy side mainly consists of
advertisers and/or ad agencies on behalf of advertisers. The sell-side consists of publishers (entities such
as New York Times) who publish ads on their websites and apps, Ad networks who aggregate advertisers
space inventory and sell-side platform.

Advertisers: Advertisers want to spend on marketing in order to grow their revenue. Some of the examples
of advertisers are Bank of America and Samsung. Nowadays they have customer data like never before
and they are trying to make digital marketing more efficient by knowing their customers better. They are
enriching their existing customer data with third party data from partners to increase their advertising
accuracy and reach. Companies are working on attribution - measuring which channel and marketing effort
has contributed how much to the sales. Overall, their attention to improving advertising performance has
drastically increased. The recent trends show that they are using rich media such as videos in their
marketing because of the reach of smartphones and internet. Also, availability of data and smart advertising

41
is creating a level playing field. Some large players are shifting all the operations of advertising from
partnerships model to propriety or in-house model so as to have marketing as strategic competency.

Agency: Agencies specializes in helping the advertisers with their creatives, managing the bidding process
in the auctions, and budget allocation. Some examples are Omnicom and Publicis. Agencies are facing
market pressure from the larger clients as they are shifting the advertising operations in-house for getting
more control and reducing costs.

DSPs (Demand Side Platforms): These are the platforms used by advertisers and media buying agencies
helping advertisers to perform real-time and automated media buying using real-time bidding. Examples
are Google Bid Manager is Adobe Media Optimizer. DSPs are central to advertising campaign
management because they facilitate campaign design, implementation, and optimization by allowing
marketers to set limits for bids and parameters for campaigns and target customers. As many DSPs interface
directly with the advertisers, the need for people handling media transactions and agencies is decreasing.
[39]

Publishers: Publishers are generally business who have an online presence through website and apps. They
have space (also called inventory) where they can serve ads to their visitors. Most publishers have a revenue
model of earning through ads served such as Facebook, Yahoo, etc. They try to generate traffic on their
websites and serve relevant ads to the target customer segments. The cost of the inventory sold by
publishers is based on the amount of customer traffic they get. They have their premium inventory/space
(the more visible space on the website that catches majority eyeballs) and remaining space which is below
the fold (website needs to be scrolled down to see it) or at the bottom of the page. Traditionally all the
inventory used to be sold by the publishers through the ad agencies who in turn would negotiate deals with
the marketing teams of the advertisers. The programmatic advertising technology has changed the way in
which the remnant inventory is sold. Though premium inventory is being negotiated traditionally today,
publishers are increasingly moving towards the programmatic technology. [39]

One of the major advantages of programmatic technology is the data it offers about the post-
advertising/post campaign performance which lets the publishers make an informed decision about how to
monetize the space more efficiently and helps advertisers use customers' behavioral data to target them
better with relevant content and customization.

Supply Side Platform (SSP): It is a platform used by publishers to sell the ad inventory on their websites
using programmatic technology and enable routing to demand side including ad exchanges and DSPs.
Some examples are Pubmatic and Google Admeld. The algorithm in SSP is designed to manage ad yield
of publisher ad inventory. SSP's function is also to ensure the best price for supplier inventory during

42
transaction or bidding. Using SSPs, publishers can control the transaction to a great extent such as setting
a minimum price for their ad space, and buyers whom they want to sell to. They can also dynamically
decide to switch between channels through which they want to sell their inventory - for example, from
direct selling in the market to Real Time Bidding. The process that publishers follow in the SSP to make a
transaction deal is to enter details of their desired criteria in the SSP interface such as the type of space,
media format, customer segments, etc. A deal ID is created and send to the DSPs or an agency/advertiser.
Some SSPs are enabling a private marketplace for some selected agencies/advertisers to provide them
priority access and first chance at the premium inventory. [39]

Industry bodies: One of the primary industry body is a global trade organization called IAB (Interactive
Advertising Bureau) consisting of approximately 650 members. It includes companies of stakeholders in
the digital advertising ecosystem. IAB helps develop standards, conducts and publishes research and best
practices on related topics. In addition, it offers training and education to its members through classes,
seminars, etc. Other such industry bodies include TAG (Trustworthy Accountability Group) and ANA
(Association of National Advertisers). They also work on topics including fraud detection, digital security,
and privacy. [39]

Specialised Industry vendors: These include third-party data providers who collect customer data from
multiple sources such as credit card companies, package the data and sell it. This data is used by companies
such as Amazon to enrich their propriety data with missing information.

DMP (Data Management Platform): This is a complex system that consolidates data from multiple sources
both offline and online by plugging into different data sources, and analyses that data. It helps unify all the
data for a particular customer to provide a single view. DMP's have interface that helps send data to
multiple channel partners so that they can activate their campaigns. Examples of DMPs are Adobe
Audience Manager and Google Doubleclick Audience Center).

Publishers, advertisers, and agencies use DMP in different ways. Publishers use the data and analyze it to
accordingly manage the inventory of space they have for offering to advertisers efficiently. When
publishers augment their various types of inventory spaces with the data of customer response to ads in
those spaces, it can improve the value of publishers inventory. One of the primary ways of analyzing the
data is to create customer segments based on their ad interactions and visits. This is done either manually
by creating groups of common interests among customers or using look-alike modeling where the
application of Machine Learning models can learn to segment and target new customers. Advertisers and
agencies use DMP to create and manage datasets for activating the campaigns from clients on different
channels. [39]

43
DMP has three step process:

Aggregate: DMP (Figure 20) helps identify most useful data from disparate sources such as ERP, CRM,

customer subscription data, and analytics systems. It also provides insight into the audiences (target
customers) by helping track data from multiple devices.

At
I I I I A 1A
lIl I I igil iI IAAII A I
IiIIAIA.
lil Il
Data Matching Activation
Multiple disparate records for The DMP's power is its ability to unify The DMP can make data usable and
the same customer myriad data points to connect all of a shareable for channel managers to
customer's data activate within their campaigns

Data Inputs Live Channels

Adobe Audience Manager
I st Party Store
Only Customer
Data Data Intake

Data 0
Nlormalization

Segmentation 0
C and scoring
Ist ~ ~ Part Str
N
0

Tag 0
Management tS
x

Ausienc
Insights No PHl
input output

CURRENT CURRNT FTUX

-

Figure 20, Representation of Adobe's Data Management Platform [40]

Segment and find target customers: First, the DMP helps identify important indicators and insights from
the gathered data. Second, it helps organize the data to create customer profiles based on a variety of
customer-related information including device, purchasing history, contextual, interest, location, etc.
Third, it helps capture the results of previous campaigns and optimize new ones. Finally, it helps generate
customer segments, turning data into a set of customer who can be targetted.

44
Action: DMP has links with marketing tools and channels, for example, some companies integrate DMP
with ad servers, email tools, and interfaces of publishers such as Facebook, Pinterest, etc. It operations of
inward and outward data flow take place in real time resulting in quick insights. Segments are updated in
the DMP after the feedback from the campaigns.

The step two above consists of data analysis that helps companies know more and granularly about their
customers and find the look-alike customers is the most fertile ground for applying Machine Learning
because of its potential to automate the task of analysis and because of the learning required with each new
customer and each new information about the existing customer that requires updates to the segmentation.
This is discussed further in the next chapter.

4.5.3 Technologies enabling advertising

Cookies: Also called HTTP cookie, it is a piece of data that the website (example, Amazon.com) sends to
the customer's web browser (example, Google Chrome webpage). After the cookie is sent, the browser can
fetch that data from the server and update the cookie. The intent behind designing cookies was to keep
track of the state of users browsing on a particular website. With the use of cookie, e-Commerce websites
can maintain the state of their customers' information from previous browsing events which buttons they
clicked on, how much was their browsing session time, their time on various pages, products they added
to cart, their information such as name, delivery address and passwords for a website, etc. There are two
types of cookies - first, a persistent cookie that has a long-term expiry date which could span months or
years and is designed to maintain a long-term state of customers' activities. Second, session cookie that is
temporary and maintains a state of browsing only till the browsing session ends. [39]

When the domain attribute of a cookie is same as the server which sets it, it is called the first-party cookie.
But websites have some images or Advertisements which could be contents of other domains. This content
from other domains when read and retrieved called third-party cookies. Generally, first-party cookies can
track activities within the same domain, whereas the third party cookies can track activities across other
domains. For example, the advertisement on Amazon.com if has third-party cookie can track customers
activities across other websites. Because of the privacy concerns, most web browsers servers are starting
to block third-party cookies. [41]

A cookie is useful because it identifies customers, stores their information about various website visits - a
compilation of which can be used for can be analyzed for identifying customer interests and also categorize
the customer into the segment and serve them relevant content or ads. It is important for the advertising
ecosystem that the cookie data is synchronized across various domains.

45
Pixels: A pixel, also called as a stage, is a piece of HTML code or JavaScript in the form of a minuscule
image that could be embedded in the webpage or in the marketing email sent to the customers. This pixel

tracks the customer's information such as their activities of websites, the information about customer's

interaction with the marketing emails or ads. Pixels from a particular sender identifies whether a cookie
from the sender was set at the customer's browser. When a customer opens an email or any interface with

a pixel, the pixel enables its sender website to place a third party cookie at the customer end. Digital
marketing activities require managing these tags for efficient information tracking, and various systems
such as Google or Adobe tag manager does this task. [39]

Ad servers: Ad servers store and deliver online ads. Other functions of ad servers are to manage campaign

traffic, optimize check ad performance, monitor and report, and campaign analysis. There are two
categories of servers - publishers' and advertisers'. Publishers are parties that publish ads on their website
such as New York Times. Publisher ad servers do campaign management and prioritization of campaigns
from several clients. Remote servers (third-party servers) distribute ads across multiple domains owner by
several independent publishers. Another type of servers called local servers are owned and operated by the

single publisher to publish ads on its own domain, providing them better control. These servers Advertisers
buy advertising space from various websites, apps, etc. Ad servers for advertisers help them manage ad
distribution across myriad media and publishers, track the purchase of media, send and monitor ads, and

report performance. [42]

Ad networks: These are brokers who aggregate the ad space from various publishers and sell this inventory
to the advertisers with a markup.

Ad exchanges: This is a marketplace where with the help of Real-Time Bidding (RTB), publishers sell the
advertising space (on their website/app) or inventory to the advertisers at a maximum price by means and
advertisers attempt to purchase the inventory at fair value with some information about the past bid prices.

The examples of ad exchanges are Google's Double Click and Microsoft Ad Exchange. [39]

4.5.4 Programmatic advertising - a closer look

There is more than one way in which the transactions take place in the space of digital advertising. Broadly,
the transactions can be categorized into fixed price and auction-based.

46
 rivl~ttkcn-ax

Open

Q2 guaranteed fixed rate

Reserved unreeved
Type of rventory

Figure 21, Programmatic transaction types [43]

Broadly, there are four types of transactions as seen in the Figure 21 based on whether the inventory is
reserved and whether the price is set. Open auctions are most common of all types of transactions and
generally can be accessed by all buyers. Invitation only transactions, on the other hand, are open to select
advertisers and publishers. In the automated guaranteed transactions, as opposed to unreserved fixed rate,
the inventory of space is set aside for the buyers who want to have a choice on where they want to show
their advertisement and want to ensure advertising space for particular campaign delivery. The process of
these deals is partially automated. It is enabled by the Open Direct protocol and is maintained by IAB
which works to not only set standards and make the communication between stakeholders but also enable
growth and innovation in programmatic advertising.

The auction-based transactions are enabled by Open Real-Time Bidding (RTB). The auctions take place
within 100 milliseconds. The complex transaction process is explained schematically in the Figure 22.

47
 Advertisers Cookje Publisher (Website)

POT T E RY consumer
B A RN ceent)
Direct Sale r
- -.. - - - - -Data Lyer--
Data Platform
gAgency Third Party Providers
Omni m ---- RealTime Siding
Trading Desk (~ 10 ms)

Dm P fomAr
vEccuen
doubed.. admdfi
RT ., Budget4'1rocketfueL echck (UblCoo RTB, Yield
Demand Side Platform Ad Exchance SuPPlY Side Platirm Mnm

Figure 22, Representation of advertisement inventory transaction process [44]

An Advertiser can buy directly from the Publisher when premium inventory is sold typically on a
guaranteed basis by the Publishers sales team. And advertiser can buy from Ad networks that are
aggregators of the ad inventory from Publishers. This is typically sold on a non-guaranteed basis. And
advertisers can also buy from Ad exchangers through DSP which has RTB capabilities and can buy
inventory with the advertiser on the Ad exchange. The trend is moving increasingly toward Real-time
bidding.

Real-Time Bidding: Publisher make the ad inventory available for advertisers to buy on the Ad exchange
and they want to maximize their price. Advertisers work with DSPs who enable them to access Ad
exchanges. DSPs bid on the inventory based on the past data analysis, and try to pay a fair price. When the
customer visits a website, before the ad is displayed, an ad call is made to the exchange by SSP. Ad
exchange hold auction for each impression of the ad. It sends bid request to each DSP to bid on that
impression. DSP sends out a bid amount and winning the bid is chosen by the ad exchange and the ad is
displayed on Publisher's website.

48
 Real-time bidding
auction

aweten
2 S8Pmaaues
abidreque t 3 &8Pevakates
theWrequ"
4 DSPsends
6 8Pevduatb p
bcirecpnsesp
6 sWProK*ef
7 O6Psends
adrnw"up
8 B3Pen~Aes
edsevmf

Figure 23, Real Time Bidding process [43]

To summarize, the major technologies underlying e-Commerce can be classified as e-Commerce website
and technologies in advertising ecosystem. Major technologies on which application of Machine Learning
will be illustrated are discussed. They include recommendation systems, product search and product detail
page. Advertising ecosystem which includes platforms including Ad exchange, Data Management
Platform, Supply Side and Demand Side Platforms, Ad server, and third party system, as well as transaction
process are discussed. Mainly the complex process of RTB is discussed in detail. With the understanding
of these technologies powering e-Commerce ecosystem, we discuss application of Machine Learning to
these systems in the next chapter.

49
 5. Applying Machine Learning

5.1 Introduction to Machine Learning [ML]

5.1.1 Description

Algorithms that enable machines to "learn without being explicitly programmed". ML models learn by
identifying patterns from experience (provided data called training data) and apply this learning to new
data. [45] ML can be applied to tasks in many areas such as face recognition, document classification,
disease diagnosis, speech translation, etc.

ML problems can be described by three parameters: task, training data and performance metric.
Following description discusses ML problem with an example [46]:

Task: Spam email detection

Training Data: Data that consists of a list of emails with label Y indicating whether an email is a spam. Y
will be either 1 or 0 depending on whether it is spam or not. The data will also have attributes of email,
say X that email can be represented by.

Suppose email E can be represented by attributes X 1 , X2 , X 3 ... Xl could stand for 'email source
unknown', X2 for 'email subject line contains $', and so on

The goal is to come up with a function Y =f(X1 , X 2, X 3 ... ) using this training data. This function results
in a decision rule which model learns from training data. When a new email is provided to the model, it
can decide based on this decision rule, whether the email is spam or not.

Decision rule could be that email is a spam, that is Y=1, if X, and X 2 and X 3

Performance metric: The metric could be to reduce probability of mistake

Fundamentally, model selection in ML for this problem is finding a set of decision rules that can be
generalized well to classify a new set of emails into a spam or not spam.

5.1.2 Classification of ML

Machine Learning can be classified on a couple of bases. First, it can be classified into following
categories on the basis of how model learns [45]:

Supervised learning: These are models that learn based on examples or feedback. The example of spam
email classification was a supervised learning task because it knows from training data which emails
could be spam based on their attributes.

50
Unsupervised learning: These models do not have feedback or information about the outcome to learn
from. These tasks are generally like partition or clustering data into similar groups or describing the
structure of the dataset.

Reinforcement learning: This model works with a feedback in the form of rewards or punishments which
is given dynamically when a machine performs a certain task. For example, a robot is required to
successfully hit a ball. When it hits, it gets rewarded and when it fails, it gets punished. The objective of
the model is to maximize its rewards.

Second, ML tasks can be categorized as follows based on the type of output (major categories mentioned,
not exhaustive):

Classification: Inputs are divided into classes

Regression: Output is continuous rather than discrete

Clustering: Output is groups which the inputs are divided into

Density estimation: output is in the form of distribution of inputs

Dimensionality reduction: When inputs have many dimensions, the output is input mapped to low-
dimensional space

5.1.3 Choosing a model

Steps for choosing a model are [47]:

1. Categorize the task by input and output

2. Find applicable and practically implementable algorithms

3. Apply all of them and compare their performance

4. Select one and optimize its parameters

Following (Figure 24) by Microsoft is a rule-of-thumb for selecting models, for people who are learning
the basics. It is compiled from contributions, experiences, and feedback of data scientists.

51
 Figure 24, ML Algorithm Cheat Sheet [48]

5.2 Machine Learning Canvas

At times, the data science teams invest considerable time in building Machine Learning models, but at the
end of the exercise and application of models, the teams do not see Returns on Investment. This framework
starts with the idea of value proposition or a business case that managers need to define before beginning
to work on Machine Learning tasks. This is a good starting point for a designing system that uses ML.

52
 The Machine Learning Canvas (vO.4) ,E
*1to45155r2
~
..............................
I
Decisions ML task Value Data Sources Collecting Data
I Propositions
H-ow are pr'edlcdonsd to kpj J oupIQ wedt
ake decisionsthat provicde Wichi taw date sources can H"o do we "e new da o W
W9p of prairie. weUse (itena end lear from (in"ol &Wd
the proposed value to lOw nd-titer? Whet ame are "y to do forth.
end-uer(s) of fh. Preictive ssems? Ordernad)?
Whobjolecives oae we serving?

|
Making Offline Features Building Models
Predictions '- Evaluation VX
hupri reproserolehra Imr dosev creelelpdate
Whren doe mos predkdlam oe er MelodoaMW eHrefls to evaluat mre eraeredfto few dale medelWth now trairto
topelo? Howlonog dowee Ieo sysiem befre deloy-wmL. date? 11w kong do weloave to
teekafte a MoartpW Al arit feafudze hahuhog w4PrdAerd0 creak. a
prarkiciour? model?

Uve Evaluation and

Monitoring

*sstm atdeyra, WArr

-mwVOUOCesn

Figure 25, Machine Learning Canvas [49]

The framework starts with defining the value proposition - the what (the goal of the system), the why (the
system is needed) and the who (is going to use the system and get impacted by the system) of the system.
Next step is to think about how the system will predict (left half of the canvas) and how it will learn(right
half of the canvas). ML task asks about the background information - what the system is trying to achieve,
what input is required and what output analyst is trying to predict, what type of ML problem it is such as
classification, regression, etc.? Next step 'decisions' is to understand how are predictions being used to
make decisions and 'making predictions' is to plan details around making predictions such as constraints,
time for making predictions, plan to evaluate the system offline prior to deployment, etc. [49]

Next, (to the right of the canvas) the factors to consider are first know how to collect data and what are the
data sources, second which features are to be used. Features are the characteristics or attributes of the inputs
such as gender, location, age, etc. of the customer. Finally, consider the details about building the ML
model. The final step is to evaluate and verify system performance.

5.3 Application of ML to Recommendation Systems

53
 Recommendation systems originated in the 1990s and have evolved since then. Today, these systems have
become more and more important for helping customers make decisions about the products or services they
want to consume as the consumption is becoming digital (for example songs consumption of Spotify, videos
on Netflix, products on eBay, etc). The researchers have been studying the use various Machine Learning
(ML) algorithms to different applications of the recommendation system in order to provide better
recommendations. There are many ML algorithms and approaches used in the recommendations systems
till date, and they are based on the logic of the recommendation system, the data about the customers and
the recommended items. Because of the sheer variations in ML algorithms, their combinations and the
number of different algorithms used in the past to design recommendation systems, it has become
challenging to choose and track applications of ML algorithms in these systems. [50]

Application discussion: Suppose that a Product Manager wants to design a recommendation system for a
PDP (Product Detail Page) of an e-Commerce website that recommends 'what other people bought' or
'what you might like' products on the page. Following is the example of recommendation canvas for
recommendation system. So the objective is to predict the product affinity. Let's say that customer C is
interested in purchasing a laptop P. Website knows that she prefers ratings of 4.5 and which all laptops
online have that rating. The problem is to find how the customer will rate laptop P. The function can be
represented as Affinity (CIP) (Affinity of customer C for product P).

ML Canvas: Recommenders

Figure 26 MLava a to L or~

re omendation system[a

b Orifte chniueon s tes

applied. The fundamental concept is to use the information about what other customer likes to predict what

54
the other customer might like. The problem can be solved by Matrix Factorization method. Customer-
Product affinity matrix can be represented with all the customers in the database on one axis and all the
products on the other axis (Figure 26). When all customers ratings for all the products are filled in the
matrix, the observation is that the matrix is very sparse. This means that very few affinities are known
explicitly as a few customers rate a few products. The problem is to predict the unknown affinities - the
ratings a customer might give to a product she has not rated (used for ranking and displaying as a
recommended item). The issues are that very small amount of data is available, and the matrix is very large.
The Machine Learning model needs to try to predict the missing values or the ratings for the product with
available only around 1% of the values for product ratings.

Input: Input will be the features/attributes of the all products P and all customers C, and the rating of
customers C for products P

Output: Rating of new customer for a product P

Model: The abstraction use is that the large matrix is projected by two small matrices X (representing the
customer) & Y (representing the product) (Figure 27). This is called as Link Factor Model. First, the model
is trained by using the data that is already available and then apply the learning to the new data. The very
first step is to run similarity using a function such as a cosine similarity to find the products that are similar
to other products for which ratings are known. Cluster algorithms can be used for these. To start with, this
helps get the initial approximate for the missing values of product ratings. Getting the customer data about
their possible interest in products they haven't rated (Pc) can also be done by tracking the real-time customer
behavior. For example, if a customer spends more time interacting with a product, her interest in the product
is inferred to be high, and sample data can be collected and plugged in the model for retaining it with real
data to obtain predictions. Next step is to rank the products with the information about the customer based
on prediction (Xc). [50]

P P

User product affinity T

User product affinity matrix x

.

Item factor
C matrix matrix

User
factor
matrix

Figure 27, User Probability Matrix [50]

55
Now, Approximate X, = X (CIP,)

Z =X.YT

Z= X.YT

PC ~ Xc .YT

pc(YT)- ~ X, .yT(yT)-1

Xc ~ Pc Y (yTyT)-1

Consequently, all the unknown values including Xc can be generated.

Next, parameters need to be included in the function in order to factor the assumption about the number of
existing data points to be used (yT), the number of iterations to train the model for, and parameter to avoid
overfitting. Finally, use Root Mean Square (RMS) error to find how well does matrix approximate the data
to evaluate the model to find which value of parameters present the best value for RMS. [50]

Improvements - For websites such as Amazon.com, the matrices are very large because of the long tail of
products. So, collapsing technique can be applied to the matrix - collapsing products by categories or by
some features. Second, currently, the matrix is two dimensional with products and customers, however,
adding the third dimension of time helps predict better because ranking for a customer changes with time.

Other ML methods: A systematic literature review of recommendation systems has been done to study the
most used applications of ML algorithms. Some ML algorithms found in the literature review can be
classified easily because they are small variations of widely used and established algorithms, but others
cannot be classified intuitively. [51] Out of the 35 sources used for the review, 26 were retained based on
the quality and clarity of approach and validity of sources. Following ML algorithms have been used in
recommendation system (listed by order of frequency of use in the retained sources).

56
 isayesian
Decision Tree 5
Matrix factorization-based 4
Neighbor-based 4
Neural Network 4
Rule Learning 4
Ensemble 3
Gradient descent-based 3
Kernel methods 3
Clustering 2
Associative classification 1
Bandit 1
Lazy learning 1
Regularization methods 1
Topic Independent Scoring Algorithm 1
Figure 28, ML algorithms for recommendation systems [51]

5.4 Application of ML to e-Commerce Search

E-Commerce search has been discussed in Chapter 3. In this section, we discuss the application of Machine
Learning to e-Commerce search. Learning to Rank technique has been widely applied to search and ranking
widely such as in the web search. We discuss the application of Learning to Rank method (specifically
Logistic Regression ML model) to this problem. Continuing the discussion on search - the problem is
mainly about combining many different attributes of the query that determine the quality of output
(accuracy of products presented as a result of the search query) to generate a ranking of products. The
objective is to rank the products by their relevance to the customer and improve accuracy and robustness
of ranking. ML model is used to optimize the weights of attributes while combining them by making use
of the customer feedback about product relevance in the form of clicks, add to basket, purchase, etc.

Application discussion: Consider for a query-product pair (Q,P), there can be many attributes Xi (Q,P)

Inputs Xi include:

- Content-based attributes such as the keywords mentioned in the query used as a basis for matching
product by application of a retrieval function (a function such as Query Light Code and PL2 that
retrieves product whose title or description matches with the search query entered by the customer

57
 looking for a particular product on the e-Commerce website). The features include various product
features as discussed in e-Commerce search section in Chapter 3.
- Behavioral features of the customer searching the product, business metrics and other features such as
product availability, etc. as discussed in Chapter 3.

Output: Probability that the customer will interact with the product presented as a result of the search
function is p(R=1 IQ,P). This function represents the hypothesis that the probability that a customer finds
a product relevant is related to the attributes factored in the inputs "through a particular form of the function
that has some parameters. These parameters can control the influence of different features on final
relevance." This hypothesis is validated experimentally. [32] [34]

The next idea is to understand the relationship between the input attributes with the output by understanding
these parameters and to understand the which attributes are more influential on the relevance (the output).
This is a Machine Learning task. It is based on the past data that is the training data which provides evidence
of which attributes are important by providing information about when customers interacted with a
particular product. Simply put, the training data will provide clues and evidence on when a customer found
a product relevant (a successful output for the search function) and what was the relation of the input
features with the output in the successful instances. The assumption here is that the products that are clicked
or purchases are better with respect to the searched query than those that are not. [32] [34]

p(R=lI Q,P)= S(XI (Q,P), X2(Q,P).....Xn(Q,P), X)

X is a set of parameters. It can be learned by optimizing the above function on the training data.

Model: A method based on regression will be used to analyze the relationship between input attributes and
the output in the training data. Many different regression-based models can be used, but here Logistic
Regression model is considered primarily because the output required is the prediction whether the product
is relevant to the customer - binary output - yes (1) or no (0). Also, this regression method is very simple
to start with and has been widely applied to this category of problems successfully.

In this case, we assume that the input attributes can be linearly combined. Let Xi be the parameter that
controls the weight of the input attributes. A higher value of X means greater weight of a respective attribute
and higher influence on the output. P is the probability of product's relevance to the customer and it will
take the value between 0 (not relevant) and 1 (relevant). The attributes could be combined simply using
Linear Regression, using the expression on the right-hand side of the equation below. However, then it
could take any value even greater than 1. So, using the log function on the left-hand side of the equation,
we convert it into probability value. [32] [34]

58
 log P(R = 1 I Q, P) = Xo+ Ii=ixoXi

1 - P(R = 1 1 QP)

Next, the function above is written in the form of the probability function. If the expression on the right-
hand side of the above equation is large, then the probability value in the equation below is high. That is

the probability of relevance of the product is high. The equation below connects the attributes of the search
with the probability of product relevance. [32] [34]

P(R=11 Q, P) = 1
1 + exp (- Xo - " i=1 XoX)

Next step is to estimate X, the parameter values using the training data in order to apply this function to the

new data - Machine Learning. To illustrate this point in simplified manner, suppose the input attributes XI,
X2, X3 have values 0.7, 0.11, 0.65 in one of the instance A of the training data in which the customer found
the product relevant (R=1), and values 0.3, 0.05, 0.4 in another instance in which she found it irrelevant
(R=0). Using the above equations, the value of parameter X can be estimated. Several such instances of the
training data will help estimate X accurately (called leaning). With the equation having the X value defined,
the machine can predict the relevance of the new instances of data. In this use case, with any new customer

search query, the model can predict the product relevance and rank products based on that. [32] [34]

Comparison of ML methods for e-Commerce search: Many regression and classification approaches have
been applied to Learning to Rank for search. LamdaMART has shown better performance empirically for
the e-Commerce search by a study that compared these algorithms.

RankNet

RankBoost

AdaRank

Random Forest

LambdaMART

Logistic Regression

SVM Classier

Figure 29, ML algorithms for e-Commerce search [52]

5.5 Application of ML to advertising ecosystem (DSPs) for prospecting [53]

59
The DSPs (Demand Side Platform), as discussed in advertising ecosystem in Chapter 3, need to decide
whether an ad is to be shown to a customer and set optimal bid amount for each ad impression for that
customer. So, it needs to find a set of target customers. Depending on the goal of a marketing campaign,
the target customer (or customer prospecting varies).

Consider a problem of targeting customers who have never interacted with a particular brand (as per the
past knowledge of the customer from their browsing history that the advertisers have by means of pixels
and cookies - ref. Chapter 3). This is a key area for applying Machine Learning because of the large amount
of granular customer behavior data, data on customers' interaction with advertisement or brand, and need
for DSPs to make real-time decisions and hundreds of decisions every second on whom to deliver ads at
what price while communicating with the RTB on the Ad exchanges.

There are several challenges that have been faced in applying Machine Learning in this area. The
dimensionality of the data is very high having millions of attributes (input variables for the model). Not all
are important or worth considering. High dimensionality results into a problem of 'cold start' for a Machine
Learning model, that is, when the training data has so many attributes but the new data on which the model
us to be applied does not have data for those attributes, the model predictions for the new marketing
campaigns are not accurate. Further, collecting training data that has all values for these millions of
attributes is beyond expensive and time-consuming. Another main challenge is a high possibility of bias in
the models. The data available for marketing campaigns have relatively very low positives or conversions
(instances of customers interacting with the advertising brand as a result of viewing the ads is very low).
So, a limited sample of training data set with very low instances of positives does not train the model
accurately and bias gets build into the model. Carefully selecting and experimenting with a randomized
sample is very costly. So, it is important to have a training data set with a correct population of positives
and negatives.

One of the recent researches has tried to address these problems with application of Machine Learning in
this area by developing a two-step model training method. In the first step, the model uses a large amount
of raw data (though biased) to deal with dimensionality and sparseness. In this step, the data is collected
from multiple sources, model learns the relationships and the important attributes. The model in step 2 uses
the learned functions from the step 1 as inputs, and combine them to produce calibrated output. This
approach is termed as transfer learning. Its aim is to apply its learning to the tasks that are different from
the task model is trained on in terms of relationships of attributes with the outcome and/or have a different
distribution of data. Despite these differences, the learning can be applied to a real task that is dissimilar as
mentioned before. This aspect of the method makes it relevant for application to targeting customers for
various marketing campaigns as each marketing campaign is different in terms of performance criteria

60
Application discussion: The goal of the model is to predict using the high dimensional customer data
whether that customer will make a purchase or convert (in general terms) within seven days of being shown
an ad. Based on this prediction, the advertiser will bid for an impression to be shown to a particular
customer.

Step 1: Model is high dimensional and so complex. Not many positives and so biased for the purpose of
learning. Positives are the actions of customer indicates actions of the customers that indicate interaction
with the advertiser such as a visit to the advertisers' website. Other browsing events are negatives.

Input 1: Anonymized URLs (converted to binary form) of pages browsed by the customers in the past. The
sample consists of both positives, that is, instances when a customer interacted with the brand and negatives,
that is general browsing without interaction. The input features X are in binary format.

Output 1: Y - 0 or 1 based on whether the browsing event was negative or positive respectively. 1 indicates
conversion or customer activity that advertiser considers important from point of view of understanding
customers interest in the brand, such as website visit, clicks, purchase, etc. These observations are based on
real data collected from past campaigns.

Model 1: For the Step 1, Logistic Regression is used for training the model. SGD (Stochastic Gradient
Descent) is used to train Logistic Regression model

Step 2: The model in step 2 consists mainly of the output of model from step 1. The output of step 1 reduces
the dimensionality. The benefit is that the model gets rid of bias because the ratio of a number of positives
to the entire population in the training sample is near to real universal data set.

Input 2: Sample of customers who have seen an ad (without being targeted) fewer positives (customers who
interacted with ads). This type of sample is taken because it is like the final target population for which the
model is trained - prospects (customers who haven't seen that ad before). Input feature X = (I, Xstepi) where
Xswepi is set of all functions f(X) trained in step 1. Xstepi - [f(X), f2 (X), f3 (X)...)

Output 2: Y - 0 or 1 based on whether customer only browsed or interacted with the advertisement
respectively with n days of advertising

The trained model from step-2 is applied to the real new data which is in the raw form to make predictions
in real time. Applying this Machine Learning technique to targeting prospects for marketing campaigns is
important. Marketers do a new campaign almost every week, and several campaigns go on simultaneously,
and the timespan for these campaigns is often short (a few days). It is prohibitively time consuming and
expensive to make Machine Learning models for each new campaign. Hence there is a new to apply transfer
learning method.

61
Other ML methods: The method of transfer learning has been discussed in previous studies and application
of following Machine Learning techniques have been demonstrated.

Hierarchical relationships within the feature set

Laplaces moothing into Poisson regression
Use of alternative outcomes in a classification model
Multitask Learning

Figure 30, ML algorithms for customer prospecting [53

Problem identified with the above techniques is that though they use transfer learning in a way that can be
applied to marketing campaigns, the practical challenge is that cross campaign data transfer or use is not
permitted due to the privacy of marketers' data. The technique of transfer learning discussed above in the
detail solves this problem as it only learns from data which is relevant to the campaign but collected from
different sources. These sources are in the form of distributions and relationships of attributes with the
target outcome.

5.6 Propensity Modelling for customer targeting

Any e-Commerce company generates data about the customers from the transactions customer do with
them or customer subscriptions on the website. So, in all, they would have data of only the customers who
previously interacted with them. This is a small number of customers, and propensity modelling helps them
to find look-alike customers (who are not their customers or who did not purchase the product for which
marketing campaign needs to run), ones who are similar to their existing customer and might be interested
in buying particular products for which marketing campaign is to be run. The propensity models can be in-
built in the DMP (discussed in the advertising ecosystem, Chapter 3).

62
 Examine purchasers Vs Non-purchasers

-iBrowsing behavior
at a class and
category level prior
to purchase

Figure 31, Representation of Propensity Model

Input: Features of purchase behavior, browse behavior and other remaining information including the
customer demographics, quote details, etc.

Output: The model should give the probability of customer buying a product or responding to marketing
campaign

Model: As this is a regression problem, and outcome required is probability, logistic regression is used. The
model considers all the input features discussed in input (independent variables) and regress them to predict
the odds of success by determining the probability.

Application description: Suppose an e-Commerce company wants to do a marketing campaign for selling
a microwave. The first step is to find the microwave buyers in last say 1 year (depending on how much data
the company has generated). Then, find the non-buyers in the same period. Do a propensity score matching
which helps match buyers with similar non-buyers (similar one to one for most features such as
demographics, etc.). So, the data has the same number of buyers and non-buyers who are similar. Also,
enrich the data with variables about their recent purchase activities that could indicate that they may buy a
microwave (such as their purchase of the refrigerator, or new household items). Run the regression model
to understand the key features that most differentiate the buyers from non-buyers. These factors or features
could be the causal factors for purchase (for example, socioeconomic status, or moving to a new house,
etc.). After identifying these important factors with the training data, improve the model, and then validate

63
with new data. The model will provide the probability distribution of customers purchase likelihood, and
company can target customers with high likelihood depending on their marketing budget.

5.7 Product Detail Page Optimization

When a customer clicks on the product icon in the search results on the e-Commerce website, the icon
redirects to the Product Detail Page (example in Figure 32) that describes the product features, price,
delivery date, and shows related pictures and videos. This page brings the customer closest to purchase.
Also, if a customer is on the product detail page, it indicates high purchase intent. Therefore, conversion at
this stage of the customer journey is critical to purchase. Similarly, product detail page is a place for
upselling, cross-selling, and reading customer reviews.

Application details: There are several details that e-Commerce product pages can entail, but it needs to
satisfy the customer with the needed information and to motivate purchase. The company needs to identify
which of the components on the page matter most to the customer for driving purchase. They could be
different for different customer segments or even for individual customers. Therefore, companies want to
use Machine Learning to optimize and go a step further, personalize the product detail pages for their
customers.

Input: Various attributes of the product page (as numbered in Figure 32) and data about customers' past
purchase behavior.

Output: Output is to predict conversion. The idea of conversion could vary, as it could include customer
actions from activities that indicate engagement to customer purchase. For the discussion, we will consider
conversion as customer's engagement with the product page. Measuring product page quality by purchase
metric will not provide accurate idea because a customer's decision to purchase may be a result of product
features or some other reasons that are not related to the product page. Product page can motivate a purchase
only by providing accurate and desired information in best possible layout or format and make the purchase
decision easier and faster. It cannot change the customer's opinion about a product. Engagement can be
measured by behaviors such as the number of clicks on the product page, number of scrolls, and the time
spends on the product page.

64
 0 0 s tanley 60-100 10-Piece Standard Fluted Screwdriver Set

-e $9.99& FREE Shipping on orders over 49 etas

&n Stock.
*Wila it nomorrow. Sept. 15? Order W"htn 6 hrs 30 WAs and choose One*Day Shipping at Checko t Detais
0006 from and sold by Amazomcnm Glftranp availab4e,
Packalp Owgttu 1
2 3 4 5

Stanley - 8ack 4 Dedate

Harderwd lips gve nonslp lt for use on even the most stripped screws
Heat treated 0"j steel btades with rust resistant nkckel-plated her, ergonormic, tn-oblar, sip-reslstat henres
Includes trond and 2-point Ph4lps. 4 slotted 2 stubby, and 2 pocket screwdtvers
Backed by EdtnseM bankd wart"l
Hardened tips give nonestp Ot for use on even the most stipped screws
V Show rnre

Rol oe sge to Zonin16ne mSO"2udfmV99

Figure 32, Product Detail Page [55]

Model: This is fundamentally a regression problem, and techniques such as Linear Regression can be used.
However, it is a practice to explore the application of various Machine Learning models and understand
which model provides the best prediction, before deciding which model to use for the training. In many
applications, Random Forest has shown most accurate predictions. Random forest is an advanced technique
which includes the characteristics of classification and regression method. It can use much more
observations than regression while capturing the variances in the dependent variables or the input features.
[55]

After comparing the predictive power of various models for a problem, next step is to improve the model.
Finding the significant features that have more influence on the outcome/ conversion need to be identified.
This could be done by adding and dropping the features and observing the change in the outcome of the
model. With the Machine Learning applications, generally, a function is not observed but learned from the
data and iterations. [53]

To summarize, the chapter opens with an introduction to ML, and proposes and discusses the use of ML
canvas while applying ML to any systems. Then, the chapter analyses and illustrates applications of ML to
e-Commerce technologies in detail. The application is structured systematically into sections namely the
problem statement, application details, input, output, model, and other possible models that can be/have
been applied. This chapter tries to include the logic behind choosing a model where possible. It mentions
other models that have been applied to the problems in the past.

65
 6. Conclusion

6.1 Thesis summary

Customer journeys are a series of steps that customers take before and after making the purchase. Purchase
journeys are becoming non-linear due to continual access to digital channels and increasing number of
touchpoints. Customers can jump between the steps, and with the exposure to competitors' advertisements,
they can quickly change their purchase decision in the favor or again a brand. Competing brands can enter
and exit the journey even in the stages right before purchase, and steal the deal. Also, as the customers are
purchasing digitally more and more, and as the number of digital touchpoints between customers and
purchasers is increasing, the noise of marketing is becoming an irritant to customers.

The thesis uses a framework to accurately predict the customer needs and help companies build loyal
relationships by targeting right prospects at the right time on the rightmost relevant touchpoint to assist
them in their purchase journeys. Sales and marketing departments need to look at the sales journey of the
customer not as discrete events at the touchpoint level, but in a connected way. The thesis illustrates, with
an example, an end-to-end view of the customer journey across the touchpoints and designing the journeys
of their customers' in order to consciously shape their decisions.

The thesis analyzes and illustrates the application of Machine Learning techniques to the systems of e-
Commerce website - Search, Recommendations and Product Detail Pages, and to the systems in the
advertising ecosystem - Data Management Platforms and Demand Side Platforms. The objective of all
these applications is to increase the e-Commerce sales and make the process efficient.

6.2 Challenges and opportunities for future work

E-Commerce websites have started doing journey analytics of the customers from the time they land on
their website to when they make the purchase or end the session. There is a fertile ground for applying
Machine Learning to predict and influence customers' next move and direct them to relevant touchpoint or
channel towards purchase. By using the past data of journeys that led to successful purchases, companies
can train models to predict the winning moves and use the information to help shape their customer's
journey. Further, they can optimize the journeys to have least number of roadblocks or exposure of
competing brands. "80% of leading marketers today agree that a capability in Machine Learning will be
critical in providing personalized experiences along the customer journey." [56]

However, the e-Commerce companies are yet to apply the end-to-end purchase journey analytics and
application of Machine Learning to end-to-end customer journeys. Tarun Thapar, Marketing Analytics
Manager at Home Depot, pointed out a practical challenge that e-Commerce players face in the process of

66
getting a complete picture of their customers' journeys. They can collect the real-time data about customer
actions in the domains of their and partners' websites, however, that is a limiting factor. During their
purchase journeys, customers move across touchpoints in multiple domains, and also across digital and
non-digital channels (Omni-channel). Third party cookies which collect data across domains are being used
less because of the privacy concerns, and companies cannot place pixels on all possible places where
potential customers visit. The challenge is about the broken journeys because of discontinuous journey data.
More sophisticated Machine Learning techniques need to be employed to predict the journeys with this
limitation in the ecosystem.

Another challenge is about the need for using real-time data for effective marketing. Abhi Yada, Founder
of ZyloTech pointed out that "the customer data is highly perishable". [57] There is a lag between when
customers express their interest in a product or brand and when companies learn about that customer and
target that customer to drive conversion. The ecosystem has many players and their data processing and
data transfer are naturally not connected. For example, an advertiser collects data from multiple sources,
makes creatives and ads with the help of agencies, connects with Ad exchanges and SSPs for placing the
advertisements, and by the time companies move through the ecosystem from beginning to the end, they
are working on stale data, and probably the customer is no more interested. Currently, companies are doing
mindless retargeting. For example, if a customer visits an e-Commerce website and clicks on a cell phone,
the sellers will keep displaying cell phone ads and recommendations on all touchpoints for next many days.
Some players in the ecosystem such as ZyloTech have started to use sophisticated Machine Learning and
Artificial Intelligence to tackle these challenges for the companies. However, this is the beginning.

67
 References

[1] http://www.nielsen.com/us/en/iinsights/iews/2017/a-look-at-the-evolving-e-commerce-I andscape.htin]

[2] https://d26a57ydsghvgx.cloudfront.net/email/documents/Fon-ester-Predictions 2016 The eCom-2.pdf

[3] https://en.ryte.com/wiki/Con suner Decision Journey

[4] Salesforce

[5] https://analytics.googleblog. con/2013/04/introducing-customer-jou1rnev-to-online.html

[6] https://wwwt .nckinsev.comn/business-functions/ma rketing-and-sales/our-inJsights/the-consumer-

decision-iourney

[7] https://go.forrester.comn/blogs/12-10-04-
buyer behavior helps b2b marketers guide the buyers journey/

[8]https://www.mckinsey.con/business-functions/marketing-and-sales/our-insights/the-new-consumer-
decision-journey

[9] https://ww w.business2communitv.co-n/marketin2/numbers-dont-lie-2016-nielsen-study-revealed-

refelTals-) 1477256#TRrVqUTL4gqFCJgb.97

[10] https://hbr.ore/2(010/11/using-customner-journev-maps-to

[11] http://customerthink.con/what exactly is touchpoint/

[12] http://deniseleevohn.con/wp-content/uploads/201 3/12?/VGBD-Download-Brand-Touchpoint-Wheel-

Workshect.pdf

[13]http://www.acadeimia.edu/10068745/ldentifying and Analvzing Touch Points for Building a [Jig

her Education Brand Journal of Marketing for Higher Education Identifying and Analyzing Touch
Points for Building a H-igher Education Brand

[14] https://www.traininSindustrv.com/articles/sales/touch-points-in-the-customer-jouniev/

[15] https://www. oogle.cor/analvtics/attribution/capabilities/

[16] https://hbr.org/20()10/12/touchpoints--ring-the-customer

[17] https ://www.mvcustomer.com/experience/engatement/how-to-use-buver-personas-to-support-

cu stome r-j ourincy- mapping

68
[18] https://www.mckinisey.com/business-functions/marketing-and-sales/our-insi hts/from-touchpoi nts-
to-jo urneys-seein -the-world-as-c ustoerncs-do#0

[19] https://en.wikipedia.org/wiki/Search-advertising

[20] https://en.wikipedia.org/wiki/Displav advertising

[21] http://www.admazing.ch/studien/VideoAdBenchmarks.pdf

[22] https://en.wikipedia.org/wiki/Mobile advertising

[23] https://www.mobileads.com/blog/best-mobile-ad-formats-sizes-display-ad-campaigns/

[24] https://en.wikipedia.ors/wikiiNative advertisin2

[25] https://goo._I/kHT7DW

[26] https://en.wikipedia.org/wiki/Recomnender system

[27] https://arxiv.org/abs/1511.05263

[28] https://www.amazon.com/

[29] https://ood/CdFyWB

[30] https://en.wikipedia.or,/wiki/Recommender system

[31] https://www.ijedr.org/papers/IJEDR 1404092.pdf

[32] https://www.coursera.ori/learn/text-retrieval/lecture/awVwS/lesson-6-8-recommender-systems-
collaborative-filtering-part-2

[33] http://www.cs.carleton.edu/cs comps/0607/recomme d/recommender/modelbased.html

[34] https://www.coursera.orJiearn/text-retrieval/lecture/QORNe/lesson-6-5-recommender-sy'stems-
con tent-based-fiItering-part- 1

[35] https://en.wikipedia.or/wiki/Tf%E2%80%93idf

[37] https://www.buvboxexpets.coi/amazos-product-rankine-algorithnim/

[38] https://www.youtube.com/watch?v=xnXI nxMM RO

[39] Book

[40] https://ww w.adobe.com/ex perience-cloud/topics/data-ianageienti-pIatform-dmp.htmil

69
[41] https://en.wikipedia.or-/wiki/H TTP cookie

[42] https://en.wikipedia.ora/wiki/Ad serving

[43] Interactive Advertising Bureau, 2013

[44] Forrester Research, E-Consultancy

[45] https://en.wikipedia.orv/wiki/Machine learning

[46] Machine Learning course slides, MIT EECS

[47] https://www.quora.cotn/How-do-vou-choose-a-machiine-learning-algorithm

[48] https://docs.m.icrosoft.com/en-us/azure/machine-learnin/studio/al.corithm-choice

[49] https://wwW.voutube.com/watch?v=Qpo5XzlUrwE

[50] https://www.youtube.com/watch?v=l2wLFYOUX8o

[51] https://arxiv.org/abs/1511.05263

[52] https://www.researchate.ne/publication/316998625 On Application of Learning to Rank for E-

Commerce Search

[53] https://Iink.sprin ger.com/article/10.10.()7/s10994-013-5375-2

[54] htt.s://azonservices.com/amazon-product-detail-pages/

[55] https://www.analyticsvidhva.coi/blog/2014/06/introduction-random-forest-simplified/

[56] https://ww w.thinkwith zooglecon/data-callery/detaiI1/machine-learnin-customer-journev/

[57] https://martechseries.com/analvtics/interview-with-abhi-vadav-founder-ceo-zylotech/

[58] https://api2cart.con/ecommerce/top-5 -product-recommendati on-engi nes-personalize-customer-

experience/

[59] httls ://icaom-.com/2014/1.0/09/netflix-spends-150-milion-on-conte nt-recommendations-every-vear/

70